Advertisement

Deep Learning–Based Nuclear Morphometry Reveals an Independent Prognostic Factor in Mantle Cell Lymphoma

  • Wen-Yu Chuang
    Affiliations
    Department of Pathology, Chang Gung Memorial Hospital and Chang Gung University, Taoyuan, Taiwan

    School of Medicine, Chang Gung University, Taoyuan, Taiwan

    Chang Gung Molecular Medicine Research Center, Chang Gung University, Taoyuan, Taiwan

    Center for Vascularized Composite Allotransplantation, Chang Gung Memorial Hospital, Taoyuan, Taiwan
    Search for articles by this author
  • Wei-Hsiang Yu
    Affiliations
    aetherAI, Co, Ltd, Taipei, Taiwan
    Search for articles by this author
  • Yen-Chen Lee
    Affiliations
    School of Medicine, Chang Gung University, Taoyuan, Taiwan
    Search for articles by this author
  • Qun-Yi Zhang
    Affiliations
    aetherAI, Co, Ltd, Taipei, Taiwan
    Search for articles by this author
  • Hung Chang
    Affiliations
    School of Medicine, Chang Gung University, Taoyuan, Taiwan

    Division of Hematology and Oncology, Department of Internal Medicine, Chang Gung Memorial Hospital, Taoyuan, Taiwan
    Search for articles by this author
  • Lee-Yung Shih
    Affiliations
    School of Medicine, Chang Gung University, Taoyuan, Taiwan

    Division of Hematology and Oncology, Department of Internal Medicine, Chang Gung Memorial Hospital, Taoyuan, Taiwan
    Search for articles by this author
  • Chi-Ju Yeh
    Affiliations
    Department of Pathology, Chang Gung Memorial Hospital and Chang Gung University, Taoyuan, Taiwan

    School of Medicine, Chang Gung University, Taoyuan, Taiwan
    Search for articles by this author
  • Samuel Mu-Tse Lin
    Affiliations
    aetherAI, Co, Ltd, Taipei, Taiwan

    Taipei American School, Taipei, Taiwan
    Search for articles by this author
  • Shang-Hung Chang
    Affiliations
    School of Medicine, Chang Gung University, Taoyuan, Taiwan

    Center for Big Data Analytics and Statistics, Chang Gung Memorial Hospital, Taoyuan, Taiwan
    Search for articles by this author
  • Shir-Hwa Ueng
    Affiliations
    Department of Pathology, Chang Gung Memorial Hospital and Chang Gung University, Taoyuan, Taiwan

    School of Medicine, Chang Gung University, Taoyuan, Taiwan

    Chang Gung Molecular Medicine Research Center, Chang Gung University, Taoyuan, Taiwan
    Search for articles by this author
  • Tong-Hong Wang
    Affiliations
    Tissue Bank, Chang Gung Memorial Hospital, Taoyuan, Taiwan
    Search for articles by this author
  • Chuen Hsueh
    Affiliations
    Department of Pathology, Chang Gung Memorial Hospital and Chang Gung University, Taoyuan, Taiwan

    School of Medicine, Chang Gung University, Taoyuan, Taiwan

    Chang Gung Molecular Medicine Research Center, Chang Gung University, Taoyuan, Taiwan
    Search for articles by this author
  • Chang-Fu Kuo
    Affiliations
    School of Medicine, Chang Gung University, Taoyuan, Taiwan

    Center for Artificial Intelligence in Medicine, Chang Gung Memorial Hospital, Taoyuan, Taiwan
    Search for articles by this author
  • Shih-Sung Chuang
    Correspondence
    Address correspondence to Chao-Yuan Yeh, M.D., aetherAI Co, Ltd, Room 907, 9F, No. 3-2, Yuanqu St., Nangang District, Taipei City 115603, Taiwan and Shih-Sung Chuang, M.D., Department of Pathology, Chi-Mei Medical Center, No. 901, Chunghwa Rd., Yongkang District, Tainan City 710402, Taiwan.
    Affiliations
    Department of Pathology, Chi-Mei Medical Center, Tainan, Taiwan
    Search for articles by this author
  • Chao-Yuan Yeh
    Correspondence
    Address correspondence to Chao-Yuan Yeh, M.D., aetherAI Co, Ltd, Room 907, 9F, No. 3-2, Yuanqu St., Nangang District, Taipei City 115603, Taiwan and Shih-Sung Chuang, M.D., Department of Pathology, Chi-Mei Medical Center, No. 901, Chunghwa Rd., Yongkang District, Tainan City 710402, Taiwan.
    Affiliations
    aetherAI, Co, Ltd, Taipei, Taiwan
    Search for articles by this author
Open AccessPublished:September 20, 2022DOI:https://doi.org/10.1016/j.ajpath.2022.08.006
      Blastoid/pleomorphic morphology is associated with short survival in mantle cell lymphoma (MCL), but its prognostic value is overridden by proliferation index Ki-67 in multivariate analysis. Herein, we developed a nuclear segmentation model using deep learning, and nuclei of tumor cells in 103 MCL cases were automatically delineated. Eight nuclear morphometric attributes, including length, width, perimeter, area, length/width ratio, circularity, irregularity, and entropy, were extracted from each nucleus. The mean, variance, skewness, and kurtosis of each attribute were calculated for each case, resulting in 32 morphometric parameters. Compared with classic MCL, 17 morphometric parameters were significantly different in blastoid/pleomorphic MCL. Using univariate analysis, 16 morphometric parameters (including 14 significantly different between classic and blastoid/pleomorphic MCL) were significant prognostic factors. Using multivariate analysis, Biologic MCL International Prognostic Index (bMIPI) risk group (P = 0.025), low skewness of nuclear irregularity (P = 0.020), and high mean of nuclear irregularity (P = 0.047) were independent adverse prognostic factors. Furthermore, a morphometric score calculated from the skewness and mean of nuclear irregularity (P = 0.0038) was an independent prognostic factor in addition to bMIPI risk group (P = 0.025), and a summed morphometric bMIPI score was useful for risk stratification of patients with MCL (P = 0.000001). Our results demonstrate, for the first time, that a nuclear morphometric score is an independent prognostic factor in MCL. It is more robust than blastoid/pleomorphic morphology and can be objectively measured.

      Graphical abstract

      Mantle cell lymphoma (MCL) is a rare type of B-cell lymphoma with a largely aggressive clinical course.
      • Swerdlow S.H.
      • Campo E.
      • Seto M.
      • Müller-Hermelink H.K.
      Most MCL cases have a CCND1 translocation, resulting in overexpression of cyclin D1 and subsequent dysregulation of cell cycle. Occasionally, MCL could lack the CCND1 translocation and cyclin D1 overexpression, but it still shares similar clinical features and gene expression profiles with cyclin D1–positive MCL.
      • Fu K.
      • Weisenburger D.D.
      • Greiner T.C.
      • Dave S.
      • Wright G.
      • Rosenwald A.
      • Chiorazzi M.
      • Iqbal J.
      • Gesk S.
      • Siebert R.
      • De Jong D.
      • Jaffe E.S.
      • Wilson W.H.
      • Delabie J.
      • Ott G.
      • Dave B.J.
      • Sanger W.G.
      • Smith L.M.
      • Rimsza L.
      • Braziel R.M.
      • Muller-Hermelink H.K.
      • Campo E.
      • Gascoyne R.D.
      • Staudt L.M.
      • Chan W.C.
      Lymphoma/Leukemia Molecular Profiling Project
      Cyclin D1-negative mantle cell lymphoma: a clinicopathologic study based on gene expression profiling.
      ,
      • Martin-Garcia D.
      • Navarro A.
      • Valdes-Mas R.
      • Clot G.
      • Gutierrez-Abril J.
      • Prieto M.
      • et al.
      CCND2 and CCND3 hijack immunoglobulin light-chain enhancers in cyclin D1(-) mantle cell lymphoma.
      Such cyclin D1–negative MCL can be identified with SOX11 positivity.
      • Chuang W.Y.
      • Chang H.
      • Chang G.J.
      • Wang T.H.
      • Chang Y.S.
      • Wang T.H.
      • Yeh C.J.
      • Ueng S.H.
      • Chien H.P.
      • Chang C.Y.
      • Wan Y.L.
      • Hsueh C.
      Pleomorphic mantle cell lymphoma morphologically mimicking diffuse large B cell lymphoma: common cyclin D1 negativity and a simple immunohistochemical algorithm to avoid the diagnostic pitfall.
      • Chuang W.Y.
      • Chang S.T.
      • Yuan C.T.
      • Chang G.J.
      • Chang H.
      • Yeh C.J.
      • Ueng S.H.
      • Kao H.W.
      • Wang T.H.
      • Wan Y.L.
      • Shih L.Y.
      • Chuang S.S.
      • Hsueh C.
      Identification of CD5/cyclin D1 double-negative pleomorphic mantle cell lymphoma: a clinicopathologic, genetic, and gene expression study.
      • Mozos A.
      • Royo C.
      • Hartmann E.
      • De Jong D.
      • Baro C.
      • Valera A.
      • Fu K.
      • Weisenburger D.D.
      • Delabie J.
      • Chuang S.S.
      • Jaffe E.S.
      • Ruiz-Marcellan C.
      • Dave S.
      • Rimsza L.
      • Braziel R.
      • Gascoyne R.D.
      • Sole F.
      • Lopez-Guillermo A.
      • Colomer D.
      • Staudt L.M.
      • Rosenwald A.
      • Ott G.
      • Jares P.
      • Campo E.
      SOX11 expression is highly specific for mantle cell lymphoma and identifies the cyclin D1-negative subtype.
      • Zeng W.
      • Fu K.
      • Quintanilla-Fend L.
      • Lim M.
      • Ondrejka S.
      • Hsi E.D.
      Cyclin D1-negative blastoid mantle cell lymphoma identified by SOX11 expression.
      Most patients with MCL present with an advanced stage III or IV disease with lymphadenopathy, hepatosplenomegaly, and bone marrow involvement.
      • Swerdlow S.H.
      • Campo E.
      • Seto M.
      • Müller-Hermelink H.K.
      The median survival of patients with MCL is about 3 to 5 years.
      • Swerdlow S.H.
      • Campo E.
      • Seto M.
      • Müller-Hermelink H.K.
      Identification of patients with poor prognosis would be helpful, because more aggressive treatment, such as hematopoietic stem cell transplant, could be considered for these patients. Mantle Cell Lymphoma International Prognostic Index (MIPI), which includes patient age, Eastern Cooperative Oncology Group performance score, serum lactate dehydrogenase level, and white blood cell count, is an important clinical prognostic factor for patients with MCL with an advanced disease.
      • Hoster E.
      • Dreyling M.
      • Klapper W.
      • Gisselbrecht C.
      • van Hoof A.
      • Kluin-Nelemans H.C.
      • Pfreundschuh M.
      • Reiser M.
      • Metzner B.
      • Einsele H.
      • Peter N.
      • Jung W.
      • Wormann B.
      • Ludwig W.D.
      • Duhrsen U.
      • Eimermacher H.
      • Wandt H.
      • Hasford J.
      • Hiddemann W.
      • Unterhalt M.
      German Low Grade Lymphoma Study Group (GLSG), European Mantle Cell Lymphoma Network
      A new prognostic index (MIPI) for patients with advanced-stage mantle cell lymphoma.
      High proliferation rate is also associated with poor prognosis in MCL.
      • Swerdlow S.H.
      • Campo E.
      • Seto M.
      • Müller-Hermelink H.K.
      Biologic MIPI (bMIPI), which combines MIPI and the proliferation index Ki-67, has even stronger prognostic value than MIPI.
      • Hoster E.
      • Dreyling M.
      • Klapper W.
      • Gisselbrecht C.
      • van Hoof A.
      • Kluin-Nelemans H.C.
      • Pfreundschuh M.
      • Reiser M.
      • Metzner B.
      • Einsele H.
      • Peter N.
      • Jung W.
      • Wormann B.
      • Ludwig W.D.
      • Duhrsen U.
      • Eimermacher H.
      • Wandt H.
      • Hasford J.
      • Hiddemann W.
      • Unterhalt M.
      German Low Grade Lymphoma Study Group (GLSG), European Mantle Cell Lymphoma Network
      A new prognostic index (MIPI) for patients with advanced-stage mantle cell lymphoma.
      Unlike most patients with MCL with extensive nodal involvement, a small subset of MCL cases present as leukemic nonnodal MCL, which usually follows an indolent clinical course.
      • Royo C.
      • Navarro A.
      • Clot G.
      • Salaverria I.
      • Gine E.
      • Jares P.
      • Colomer D.
      • Wiestner A.
      • Wilson W.H.
      • Vegliante M.C.
      • Fernandez V.
      • Hartmann E.M.
      • Trim N.
      • Erber W.N.
      • Swerdlow S.H.
      • Klapper W.
      • Dyer M.J.
      • Vargas-Pabon M.
      • Ott G.
      • Rosenwald A.
      • Siebert R.
      • Lopez-Guillermo A.
      • Campo E.
      • Bea S.
      Non-nodal type of mantle cell lymphoma is a specific biological and clinical subgroup of the disease.
      MCL is usually composed of monomorphic small- to medium-sized lymphoid cells with irregular nuclei (classic MCL).
      • Swerdlow S.H.
      • Campo E.
      • Seto M.
      • Müller-Hermelink H.K.
      A minor subgroup of MCL cases is composed of larger lymphoid cells with blastoid or pleomorphic nuclei (blastoid or pleomorphic variant), and these tumors have an even more aggressive biologic behavior.
      • Swerdlow S.H.
      • Campo E.
      • Seto M.
      • Müller-Hermelink H.K.
      Blastoid MCL is composed of medium-sized lymphoid cells with fine chromatin and small nucleoli, morphologically similar to lymphoblasts. Pleomorphic MCL is composed of medium-sized to large lymphoid cells with pleomorphic nuclei and more prominent nucleoli. Although blastoid/pleomorphic morphology is associated with poor prognosis, it has been shown that the prognostic value was overridden by proliferation index Ki-67 in multivariate analysis.
      • Hoster E.
      • Rosenwald A.
      • Berger F.
      • Bernd H.W.
      • Hartmann S.
      • Loddenkemper C.
      • Barth T.F.
      • Brousse N.
      • Pileri S.
      • Rymkiewicz G.
      • Kodet R.
      • Stilgenbauer S.
      • Forstpointner R.
      • Thieblemont C.
      • Hallek M.
      • Coiffier B.
      • Vehling-Kaiser U.
      • Bouabdallah R.
      • Kanz L.
      • Pfreundschuh M.
      • Schmidt C.
      • Ribrag V.
      • Hiddemann W.
      • Unterhalt M.
      • Kluin-Nelemans J.C.
      • Hermine O.
      • Dreyling M.H.
      • Klapper W.
      Prognostic value of Ki-67 index, cytology, and growth pattern in mantle-cell lymphoma: results from randomized trials of the european mantle cell lymphoma network.
      Despite the apparent correlation between morphology and prognosis, the prognostic significance of morphometric parameters in MCL has not been evaluated to date.
      In this study, we developed a deep learning algorithm to automatically delineate nuclear contours of MCL tumor cells. Morphometric parameters were extracted and calculated, and their prognostic significance was evaluated.

      Materials and Methods

      Case Selection

      The overview of the study design is shown in Figure 1. A total of 103 MCL cases diagnosed between 2002 and 2019 were retrieved from the archives of Departments of Pathology of two medical centers in Taiwan (Chang Gung Memorial Hospital at Taoyuan and Chi-Mei Medical Center at Tainan). The pathology slides were reviewed by two senior hematopathologists (W.-Y.C. and S.-S.C.) to confirm the diagnosis. Cases with less than three high-quality regions were excluded. A high-quality region was defined as a square of 0.25 × 0.25 mm, in which >95% of the nucleated cells were tumor cells and no prominent artifacts, such as crushing, tissue folding, areas out of focus, or air bubbles, were found. For difficult cases, immunostained slides were used to identify areas with a high tumor cell percentage. All specimens of biopsy or resection from either lymph nodes or extranodal sites were obtained before treatment. All cyclin D1–negative cases were positive for SOX11 (clone MRQ-58).
      • Chuang W.Y.
      • Chang H.
      • Chang G.J.
      • Wang T.H.
      • Chang Y.S.
      • Wang T.H.
      • Yeh C.J.
      • Ueng S.H.
      • Chien H.P.
      • Chang C.Y.
      • Wan Y.L.
      • Hsueh C.
      Pleomorphic mantle cell lymphoma morphologically mimicking diffuse large B cell lymphoma: common cyclin D1 negativity and a simple immunohistochemical algorithm to avoid the diagnostic pitfall.
      • Chuang W.Y.
      • Chang S.T.
      • Yuan C.T.
      • Chang G.J.
      • Chang H.
      • Yeh C.J.
      • Ueng S.H.
      • Kao H.W.
      • Wang T.H.
      • Wan Y.L.
      • Shih L.Y.
      • Chuang S.S.
      • Hsueh C.
      Identification of CD5/cyclin D1 double-negative pleomorphic mantle cell lymphoma: a clinicopathologic, genetic, and gene expression study.
      • Mozos A.
      • Royo C.
      • Hartmann E.
      • De Jong D.
      • Baro C.
      • Valera A.
      • Fu K.
      • Weisenburger D.D.
      • Delabie J.
      • Chuang S.S.
      • Jaffe E.S.
      • Ruiz-Marcellan C.
      • Dave S.
      • Rimsza L.
      • Braziel R.
      • Gascoyne R.D.
      • Sole F.
      • Lopez-Guillermo A.
      • Colomer D.
      • Staudt L.M.
      • Rosenwald A.
      • Ott G.
      • Jares P.
      • Campo E.
      SOX11 expression is highly specific for mantle cell lymphoma and identifies the cyclin D1-negative subtype.
      • Zeng W.
      • Fu K.
      • Quintanilla-Fend L.
      • Lim M.
      • Ondrejka S.
      • Hsi E.D.
      Cyclin D1-negative blastoid mantle cell lymphoma identified by SOX11 expression.
      Cases of leukemic nonnodal MCL were excluded because of their unique indolent clinical behavior.
      • Royo C.
      • Navarro A.
      • Clot G.
      • Salaverria I.
      • Gine E.
      • Jares P.
      • Colomer D.
      • Wiestner A.
      • Wilson W.H.
      • Vegliante M.C.
      • Fernandez V.
      • Hartmann E.M.
      • Trim N.
      • Erber W.N.
      • Swerdlow S.H.
      • Klapper W.
      • Dyer M.J.
      • Vargas-Pabon M.
      • Ott G.
      • Rosenwald A.
      • Siebert R.
      • Lopez-Guillermo A.
      • Campo E.
      • Bea S.
      Non-nodal type of mantle cell lymphoma is a specific biological and clinical subgroup of the disease.
      All cases were restaged according to the Lugano classification.
      • Cheson B.D.
      • Fisher R.I.
      • Barrington S.F.
      • Cavalli F.
      • Schwartz L.H.
      • Zucca E.
      • Lister T.A.
      • Alliance A.L.
      Lymphoma G, Eastern Cooperative Oncology G, European Mantle Cell Lymphoma C, Italian Lymphoma F, European Organisation for R, Treatment of Cancer/Dutch Hemato-Oncology G, Grupo Espanol de Medula O, German High-Grade Lymphoma Study G, German Hodgkin's Study G, Japanese Lymphorra Study G, Lymphoma Study A, Group NCT, Nordic Lymphoma Study G, Southwest Oncology G, United Kingdom National Cancer Research I
      Recommendations for initial evaluation, staging, and response assessment of Hodgkin and non-Hodgkin lymphoma: the Lugano classification.
      For each case, one routine section (3 μm thick) of formalin-fixed, paraffin-embedded tissue with hematoxylin and eosin stain was used for digitization. Whole-slide high-resolution digital images were produced using a NanoZoomer S360 digital slide scanner (Hamamatsu Photonics, Hamamatsu, Japan) with a 40× objective mode. Three high-quality regions per slide were selected by a senior hematopathologist (W.-Y.C.). This study had been approved by the Institutional Review Board of Chang Gung Medical Foundation (Institutional Review Board numbers 201902130B0 and 202000483B0).
      Figure thumbnail gr1
      Figure 1Overview of the study design. MCL, mantle cell lymphoma; ROI, region of interest.

      Computer Hardware and Software

      We conducted our experiments on a customized server with an NVIDIA QUADRO RTX 8000 graphics processing unit. The instance detection module and feature extracting algorithms were implemented with Python 3.7 and PyTorch 1.7 on a Linux platform. The statistical analyses were performed using the IBM SPSS Statistics 20.0 and R language on a Windows platform.

      Nuclear Detection Model

      Among the 309 selected high-quality regions, 66 regions of interest (ROIs) with a size of 132 × 132 μm were randomly sampled from 66 different cases for training (30 cases), validation (6 cases), and testing (30 cases) of the nuclear detection model. The case numbers of different morphologic subtypes were balanced in each data set (Table 1). All nuclear contours of tumor cells in the 66 ROIs (10,459 nuclei) were manually annotated under the supervision of a senior hematopathologist (W.-Y.C.) using a free-hand contouring tool on aetherSlide Digital Pathology System (aetherAI, Taipei, Taiwan). For the testing set, all nuclear contours of nontumor cells in the 30 ROIs (347 nuclei) were manually annotated by a senior hematopathologist (W.-Y.C.). The numbers of cases, ROIs, and annotated cells of each morphologic variant in each data set are shown in Table 1.
      Table 1The Details of Cases Used for Training, Validation, and Testing of the Nuclear Segmentation Model and for Morphometric Analysis
      Data setMorphologic variant
      ClassicBlastoidPleomorphic
      Training set21 Cases5 Cases4 Cases
      21 ROIs5 ROIs4 ROIs
      3296 Tumor cells732 Tumor cells438 Tumor cells
      Validation set4 Cases1 Case1 Case
      4 ROIs1 ROI1 ROI
      722 Tumor cells160 Tumor cells106 Tumor cells
      Testing set21 Cases5 Cases4 Cases
      21 ROIs5 ROIs4 ROIs
      3784 Tumor cells671 Tumor cells550 Tumor cells
      231 Nontumor cells79 Nontumor cells37 Nontumor cells
      Total cases for morphometric analysis77 Cases16 Cases10 Cases
      231 ROIs
      All three ROIs of each case were used for morphometric analysis.
      48 ROIs
      All three ROIs of each case were used for morphometric analysis.
      30 ROIs
      All three ROIs of each case were used for morphometric analysis.
      ROI, region of interest.
      All three ROIs of each case were used for morphometric analysis.
      A two-stage instance segmentation model was employed to detect nuclei in the images. Our model was implemented using MMDetection,
      • Chen K.
      • Wang J.
      • Pang J.
      • Cao Y.
      • Xiong Y.
      • Li X.
      • Sun S.
      MMDetection: open MMLab detection toolbox and benchmark.
      an open-source software for object detection and instance segmentation, and trained with COCO
      • Lin T.Y.
      • Maire M.
      • Belongie S.
      • Bourdev L.
      • Girshick R.
      • Hays J.
      • Perona P.
      • Microsoft C.O.C.O.
      Common objects in context.
      formatted cell annotations. To be specific, a hybrid task cascade region proposal convolutional neural network, or HTC-RCNN,
      • Chen K.
      • Pang J.
      • Wang J.
      • Xiong Y.
      • Li X.
      • Sun S.
      • Feng W.
      • Liu Z.
      • Shi J.
      • Ouyang W.
      • Loy C.C.
      • Lin D.
      Hybrid task cascade for instance segmentation. 2019 IEEE/CVF Conference on computer vision and pattern recognition (CVPR).
      with a ResNet50
      • He K.
      • Zhang X.
      • Ren S.
      • Sun J.
      Deep residual learning for image recognition. 2016 IEEE Conference on computer vision and pattern recognition (CVPR).
      backbone was trained to segment the nuclear contour of each tumor cell in an image. Images were randomly augmented on the fly during the training phase to increase the data variability. The applied data augmentation methods included random translation, random scaling, random rotation, random horizontal/vertical flipping, random color jittering, and random gaussian blurring. The increased diversity of training images is known to make the model more generalizable and robust.
      • Shorten C.
      • Khoshgoftaar T.M.
      A survey on image data augmentation for deep learning.
      The model was trained with a stochastic gradient descent optimizer, a learning rate of 0.001, and a batch size of 16 for 1200 epochs. Nonmaximum suppression technique was employed to remove overlapping nuclei of model predictions. It selected one of multiple overlapping instances that have an intersection over union (IoU) ≥ 0.5 by keeping the most confident instance.
      The performance of the nuclear detection model was evaluated by mean average precision. In brief, an IoU was calculated for each predicted bounding box to assess the extent of overlapping with a ground truth bounding box. A prediction was considered correct if the IoU was at least 0.5. The precision and recall of nuclear prediction were calculated from the highest prediction score object to the lowest one, ranging from 1 to 0 iteratively. The average precision of each object class was then calculated by summing up the area under the precision-recall curve. Finally, the mean average precision can be derived by averaging the average precision of different object classes. Because we segmented only one object class (namely, tumor cell nucleus), the mean average precision is the same as the average precision. A bootstrapping method was employed to estimate the 95% CI of mean average precision. In brief, the data set was resampled through a sampling-with-replacement manner, followed by evaluating the mean average precision for 1000 times. The lower bound and the upper bound of the target statistics were computationally derived by taking the 2.5 and 97.5 percentile of the distribution, respectively. Mean IoU, mean Sørensen-Dice coefficient, and average aggregated Jaccard index were calculated to evaluate the similarity between segmented nuclei and annotated nuclei.

      Feature Extraction Procedure

      For each case, an ROI of 132 × 132 μm was randomly cropped from each high-quality region. The three ROIs of each case were analyzed by our nuclear detection model, and eight nuclear morphometric attributes related to nuclear size (length, width, perimeter, and area), shape (length/width ratio, circularity, and irregularity), and texture (entropy) were extracted from each detected nucleus. The definition of each morphometric attribute is listed below:
      • i). Nuclear length: length of the longest axis of the nucleus.
      • ii). Width: length of the axis orthogonal to the longest axis of the nucleus.
      • iii). Perimeter: length of the nuclear boundary.
      • iv). Area: area within the nuclear boundary.
      • v). Length/width ratio: ratio of the nuclear length/the nuclear width.
      • vi). Circularity: ratio of the nuclear area/the area of a circle with a diameter of the nuclear length.
      • vii). Irregularity: variance of the distance from the nuclear center to vertices of the nuclear boundary.
      • viii). Entropy: randomness of the intensity of pixels within the nuclear boundary.
      The four statistical moments, including mean, variance, skewness, and kurtosis, of each attribute were calculated across all three ROIs for each case, resulting in a total of 32 morphometric parameters. The statistical moments are used to describe different characteristics of the probability density function of a random variable. The first moment, or mean, is the expected value of a random variable. The second moment, or variance, is the expected squared difference of a random variable from its mean. The third moment, or skewness, is a measure of the asymmetry of the probability distribution. The fourth moment, or kurtosis, is a measure of the heaviness of tails of the probability distribution.

      Statistical Analysis

      Differences between categorical data were assessed by χ2 test, and Yates' correction was performed when the expected frequency was less than five. Continuous parameters were compared using t-test. The cutoff value of each morphometric parameter with the highest survival influence was determined by a free R-based software Evaluate Cutpoints
      • Ogluszka M.
      • Orzechowska M.
      • Jedroszka D.
      • Witas P.
      • Bednarek A.K.
      Evaluate cutpoints: adaptable continuous data distribution system for determining survival in Kaplan-Meier estimator.
      (http://wnbikp.umed.lodz.pl/Evaluate-Cutpoints, last accessed July 17, 2022) using the cutp algorithm. In brief, a Cox proportional hazards model was used to calculate the influence of a parameter on survival. An optimal cutoff value was then determined statistically by a log-rank test through comparing the test statistic with a brownian bridge distribution. A value larger than the cutoff value was considered high. Overall survival was analyzed by the Kaplan-Meier method and compared by log-rank tests. The influence of parameters on overall survival was analyzed using univariate or multivariate Cox regression. A P < 0.05 was considered statistically significant. To further examine the robustness of each morphometric parameter on prognostic effect, a bootstrapping was performed with the number of repetitions set to 1000.

      Results

      Clinicopathologic Features

      The clinicopathologic features of our MCL cases and their influence on overall survival are listed in Table 2. The age at diagnosis ranged from 33 to 96 years, with a median of 64 years. There was prominent male predominance (83.5%), and 88.8% of cases had an advanced stage III or IV disease. The induction therapy used included CHOP (cyclophosphamide, doxorubicin, vincristine, and prednisolone) (n = 19), R (rituximab)-CHOP (n = 16), COP (cyclophosphamide, vincristine, and prednisolone) (n = 13), R-COP (n = 7), VR-CAP (bortezomib, rituximab, cyclophosphamide, doxorubicin, and prednisolone) (n = 7), hyper-CVAD (cyclophosphamide, vincristine, doxorubicin, and dexamethasone) (n = 4), R-hyper-CVAD (n = 4), BR (bendamustine and rituximab) (n = 2), CEOP (cyclophosphamide, epirubicin, vincristine, and prednisolone) (n = 2), R-CEOP (n = 2), and other regimens (n = 7). Ten patients were not eligible for treatment, and the information of treatment was not available for another 10 patients. Using univariate analysis, age >60 years at diagnosis (P = 0.0026), stage III or IV disease (P = 0.018), presence of B symptoms (P = 0.011), elevated serum lactate dehydrogenase level (P = 0.0079), Eastern Cooperative Oncology Group score of more than one (P = 0.030), blastoid/pleomorphic morphology (P = 0.0085), and proliferation index Ki-67 ≥ 30% (P = 0.0070) were significant adverse prognostic factors. In addition, bMIPI risk group had stronger influence on overall survival (P = 0.000099) compared with MIPI risk group (P = 0.00037). The survival curves of patients stratified by clinicopathologic features with significant prognostic value are shown in Figure 2, A–I .
      Table 2Clinicopathologic Features and Their Influence on Overall Survival by Univariate Analysis
      ParameterN (%)HR (95% CI)P value
      Age at diagnosis, years
      Range, 33 to 96 (median, 64) years
       ≤6039 (37.9)1
       >6064 (62.1)2.29 (1.33–3.92)0.0026
      Sex
       Female17 (16.5)1
       Male86 (83.5)1.60 (0.77–3.36)0.21
      Stage
      Some cases excluded because of incomplete clinical data.
       I or II11 (11.2)1
       III or IV87 (88.8)3.03 (1.21–7.59)0.018
      B symptoms
      Some cases excluded because of incomplete clinical data.
       No66 (76.7)1
       Yes20 (23.3)2.21 (1.20–4.07)0.011
      Serum LDH
      Some cases excluded because of incomplete clinical data.
       Normal52 (55.3)1
       Elevated42 (44.7)2.00 (1.20–3.34)0.0079
      ECOG score
      Some cases excluded because of incomplete clinical data.
       0 or 173 (81.1)1
       >117 (18.9)2.00 (1.07–3.74)0.030
      Extranodal site
      Some cases excluded because of incomplete clinical data.
       0 or 134 (37.0)1
       >158 (63.0)1.05 (0.63–1.76)0.86
      BM/PB involvement
      Some cases excluded because of incomplete clinical data.
       No33 (37.9)1
       Yes54 (62.1)1.41 (0.82–2.42)0.21
      MIPI risk group
      Some cases excluded because of incomplete clinical data.
       Low27 (31.0)1.79 (1.30–2.48)0.00037
       Intermediate20 (23.0)
       High40 (46.0)
      Morphologic variant
       Classic77 (74.8)1
       Blastoid/pleomorphic26 (25.2)2.06 (1.20–3.53)0.0085
      Cyclin D1
       Negative6 (5.8)1
       Positive97 (94.2)0.51 (0.18–1.42)0.20
      Ki-67, %
       <3063 (61.2)1
       ≥3040 (38.8)1.99 (1.21–3.29)0.0070
      bMIPI risk group
      Some cases excluded because of incomplete clinical data.
       Low17 (19.6)2.15 (1.46–3.16)0.000099
       Intermediate25 (28.7)
       High45 (51.7)
      P < 0.05 shown in bold.
      bMIPI, Biologic MIPI; BM/PB, bone marrow/peripheral blood; ECOG, Eastern Cooperative Oncology Group; HR, hazard ratio; LDH, lactate dehydrogenase; MIPI, Mantle Cell Lymphoma International Prognostic Index.
      Some cases excluded because of incomplete clinical data.
      Figure thumbnail gr2
      Figure 2Survival curves of patients stratified by clinicopathologic features (AI), morphometric score (J), and morphometric Biologic Mantle Cell Lymphoma International Prognostic Index (bMIPI) score (K and L). The asterisk indicates that some cases were excluded because of incomplete clinical data. ECOG, Eastern Cooperative Oncology Group; LDH, lactate dehydrogenase; MIPI, Mantle Cell Lymphoma International Prognostic Index.

      Nuclear Detection and Feature Extraction

      The learning curves and precision-recall curves of our nuclear detection model are shown in Figure 3, A and B, respectively. The mean average precision was 0.887 (95% CI, 0.845–0.892) and 0.836 (95% CI, 0.832–0.840) for the validation and testing set, respectively. Our algorithm achieved a precision of 0.909 (95% CI, 0.891–0.924) with a recall of 0.835 (95% CI, 0.817–0.857) in the testing set. Regarding different morphologic variants, the mean average precision was 0.837 (95% CI, 0.825–0.860), 0.875 (95% CI, 0.829–0.929), and 0.799 (95% CI, 0.736–0.844) for classic, blastoid, and pleomorphic MCL, respectively (Figure 3C). The mean IoU was 0.686, 0.680, and 0.680 for classic, blastoid, and pleomorphic MCL, respectively. The mean Sørensen-Dice coefficient was 0.813, 0.801, and 0.810 for classic, blastoid, and pleomorphic MCL, respectively. The average aggregated Jaccard index of predicted nuclei was 0.696, 0.696, and 0.685 for classic, blastoid, and pleomorphic MCL, respectively. Examples of automatic nuclear detection and contouring of MCL cases with classic, blastoid, and pleomorphic morphology are demonstrated in Figure 3D. Among all predicted nuclei in the testing set, the mean proportion of nontumor cell nuclei was 0.85% (0.93% in classic MCL; 0.89% in blastoid MCL; and 0.38% in pleomorphic MCL). The nuclear segmentation results of representative ROIs of all cases can be found in Supplemental Figure S1. The mean ± SD nuclear length, width, perimeter, area, length/width ratio, circularity, irregularity, and entropy of all cases were 8.440 ± 0.898 μm, 6.587 ± 0.710 μm, 24.900 ± 2.677 μm, 45.125 ± 9.739 μm2, 1.293 ± 0.032, 0.759 ± 0.015, 0.173 ± 0.044 μm2, and 5.752 ± 0.228, respectively.
      Figure thumbnail gr3
      Figure 3Learning curves and precision-recall curves of our nuclear detection model and examples of automatic nuclear segmentation. A: The learning curves showed gradual decrease of loss and increase of average precision. B: The model achieved a precision of 0.909 with a recall of 0.835 in the testing set. C: The mean average precision was 0.837, 0.875, and 0.799 for classic, blastoid, and pleomorphic mantle cell lymphoma (MCL), respectively. D: Our model performed well in automatic contouring (yellow closed lines) of tumor cell nuclei in classic, blastoid, and pleomorphic MCL.

      Comparison of Morphometric Parameters between Morphologic Variants

      Comparison of the 32 morphologic features between classic MCL and blastoid or pleomorphic MCL is shown in Table 3. Compared with classic MCL, 17 morphometric parameters were significantly different in blastoid/pleomorphic MCL. These different parameters were related to nuclear size (length, width, perimeter, and area) and shape (irregularity), and blastoid or pleomorphic MCL had higher mean, higher variance, lower skewness, and/or lower kurtosis compared with classic MCL. We demonstrated that the aggressive morphologic variants diagnosed by hematopathologists truly had different objective morphometric features.
      Table 3Comparison of Nuclear Morphometric Parameters between Classic and Blastoid or Pleomorphic MCL
      Nuclear morphometric parameterClassic MCL (n = 77)Blastoid MCL (n = 16)P value (blastoid versus classic MCL)Pleomorphic MCL (n = 10)P value (pleomorphic versus classic MCL)
      Length
       Mean, μm8.194 ± 0.8348.967 ± 0.6370.000739.496 ± 0.5750.000007
       Variance, μm22.031 ± 0.6963.012 ± 0.8700.0000043.794 ± 0.958<0.000001
       Skewness0.632 ± 0.3280.320 ± 0.2130.0000370.419 ± 0.3310.056
       Kurtosis4.059 ± 1.0433.122 ± 0.5020.0000023.377 ± 0.6870.048
      Width
       Mean, μm6.401 ± 0.6606.946 ± 0.4950.00247.448 ± 0.5210.000006
       Variance, μm21.062 ± 0.5221.720 ± 0.5870.0000212.445 ± 0.729<0.000001
       Skewness0.285 ± 0.3260.247 ± 0.3020.670.167 ± 0.2280.27
       Kurtosis3.902 ± 0.8073.538 ± 0.8170.113.333 ± 0.5790.034
      Perimeter
       Mean, μm24.17 ± 2.4926.39 ± 1.880.001128.10 ± 1.790.000006
       Variance, μm214.62 ± 6.2423.28 ± 7.670.00000531.34 ± 8.57<0.000001
       Skewness0.345 ± 0.3310.135 ± 0.2420.0180.192 ± 0.2530.16
       Kurtosis3.794 ± 0.8463.162 ± 0.5340.00523.164 ± 0.5140.024
      Area
       Mean, μm242.42 ± 8.7850.42 ± 7.250.0009757.49 ± 7.410.000001
       Variance, μm4181.9 ± 117.7329.5 ± 150.80.000037514.3 ± 177.1<0.000001
       Skewness0.761 ± 0.3580.633 ± 0.3090.190.720 ± 0.2410.73
       Kurtosis4.694 ± 1.5163.933 ± 1.2020.0633.903 ± 0.8150.11
      Length/width ratio
       Mean1.291 ± 0.0321.304 ± 0.0270.141.293 ± 0.3620.88
       Variance0.038 ± 0.0080.040 ± 0.0060.260.043 ± 0.0150.30
       Skewness1.389 ± 0.2991.297 ± 0.3230.271.600 ± 0.6450.33
       Kurtosis6.108 ± 2.1385.647 ± 2.1290.437.449 ± 4.9160.42
      Circularity
       Mean0.760 ± 0.0150.753 ± 0.0140.0960.758 ± 0.0180.70
       Variance0.008 ± 0.0010.009 ± 0.0010.0830.010 ± 0.0020.067
       Skewness–0.703 ± 0.192–0.662 ± 0.2090.45–0.864 ± 0.2810.11
       Kurtosis3.455 ± 0.5023.324 ± 0.5810.363.792 ± 0.8350.24
      Irregularity
       Mean, μm20.161 ± 0.0400.201 ± 0.0410.000620.216 ± 0.0400.00013
       Variance, μm40.031 ± 0.0150.042 ± 0.0150.00880.052 ± 0.0170.00011
       Skewness2.960 ± 0.8892.373 ± 0.5190.0132.532 ± 0.7730.15
       Kurtosis17.34 ± 12.0311.14 ± 5.130.04712.63 ± 6.850.23
      Entropy
       Mean5.758 ± 0.2385.707 ± 0.1650.425.777 ± 0.2430.82
       Variance0.082 ± 0.0230.080 ± 0.0140.620.088 ± 0.0220.42
       Skewness–0.082 ± 0.203–0.044 ± 0.1740.49–0.043 ± 0.2130.57
       Kurtosis3.281 ± 0.5883.270 ± 0.2080.943.128 ± 0.2840.42
      P < 0.05 shown in bold.
      The values of morphometric parameters were mean ± SD.
      MCL, mantle cell lymphoma.

      Correlation of Morphometric Parameters with Overall Survival

      Using univariate analysis, the influence of nuclear morphometric parameters on overall survival is listed in Table 4. Sixteen of the 32 morphometric parameters (including 14 significantly different between classic and blastoid/pleomorphic MCL) had significant influence on survival. Of note, all 16 morphometric parameters remained significant prognostic factors after bootstrapping for 1000 times, confirming the robustness of prognostic effect. These morphometric parameters with prognostic significance were related to nuclear size (length, width, perimeter, and area) and shape (length/width ratio and irregularity), and short survival was correlated with higher mean, higher variance, lower skewness, and/or lower kurtosis. The distribution of each nuclear morphometric parameter with a cutoff value determined by the software Evaluate Cutpoints
      • Ogluszka M.
      • Orzechowska M.
      • Jedroszka D.
      • Witas P.
      • Bednarek A.K.
      Evaluate cutpoints: adaptable continuous data distribution system for determining survival in Kaplan-Meier estimator.
      and the survival curves stratified by each morphometric parameter using the cutoff value are shown in Supplemental Figure S2.
      Table 4Nuclear Morphometric Parameters and Their Influence on Overall Survival by Univariate Analysis
      Nuclear morphometric parameterCutoff valueHR (95% CI) for value > cutoff valueP valueHR (95% CI) (bootstrapping 1000 times)P value (bootstrapping 1000 times)
      Length
       Mean, μm8.0381.80 (1.06–3.06)0.0311.79 (1.15–2.93)0.034
       Variance, μm21.6262.15 (1.19–3.88)0.0112.14 (1.31–3.88)0.013
       Skewness0.71120.50 (0.29–0.87)0.0140.49 (0.27–0.81)0.013
       Kurtosis4.1700.56 (0.31–1.01)0.0550.55 (0.30–1.00)0.051
      Width
       Mean, μm6.2111.77 (1.05–3.00)0.0321.82 (1.13–2.92)0.027
       Variance, μm20.78601.98 (1.13–3.48)0.0182.02 (1.22–3.65)0.015
       Skewness0.48141.31 (0.75–2.27)0.341.34 (0.76–2.36)0.31
       Kurtosis3.2740.72 (0.42–1.22)0.220.70 (0.36–1.28)0.20
      Perimeter
       Mean, μm23.411.82 (1.07–3.10)0.0271.84 (1.15–3.17)0.026
       Variance, μm212.071.98 (1.14–3.44)0.0152.01 (1.22–3.53)0.014
       Skewness0.54710.46 (0.23–0.90)0.0230.45 (0.22–0.82)0.023
       Kurtosis3.8570.56 (0.33–0.96)0.0360.57 (0.31–0.93)0.043
      Area
       Mean, μm237.541.81 (1.05–3.11)0.0321.83 (1.12–3.22)0.030
       Variance, μm4118.12.10 (1.18–3.73)0.0122.11 (1.27–3.95)0.012
       Skewness0.85120.65 (0.39–1.08)0.0960.64 (0.38–1.04)0.088
       Kurtosis4.7380.75 (0.44–1.28)0.290.76 (0.44–1.27)0.33
      Length/width ratio
       Mean1.2951.69 (1.03–2.75)0.0361.70 (0.99–2.88)0.037
       Variance0.030771.47 (0.74–2.90)0.271.44 (0.69–3.52)0.30
       Skewness1.4140.66 (0.40–1.10)0.110.66 (0.35–1.11)0.11
       Kurtosis5.2110.61 (0.37–0.99)0.0450.60 (0.35–0.96)0.045
      Circularity
       Mean0.76860.58 (0.32–1.05)0.0730.57 (0.28–1.07)0.070
       Variance0.0082861.49 (0.92–2.41)0.111.50 (0.90–2.57)0.10
       Skewness–0.83971.42 (0.79–2.53)0.241.44 (0.74–2.88)0.23
       Kurtosis3.3820.80 (0.49–1.29)0.350.79 (0.48–1.33)0.36
      Irregularity
       Mean, μm20.13042.63 (1.34–5.18)0.00502.64 (1.55–4.98)0.0051
       Variance, μm40.027461.58 (0.96–2.60)0.0731.58 (0.99–2.73)0.076
       Skewness3.7290.14 (0.03–0.57)0.00610.13 (0.01–0.40)0.0048
       Kurtosis12.900.47 (0.28–0.77)0.00290.45 (0.25–0.77)0.0024
      Entropy
       Mean5.9130.75 (0.41–1.38)0.350.75 (0.33–1.45)0.37
       Variance0.068661.34 (0.78–2.31)0.291.38 (0.77–2.64)0.25
       Skewness0.19780.38 (0.12–1.21)0.100.37 (0.01–0.84)0.090
       Kurtosis3.1880.65 (0.40–1.07)0.0880.65 (0.39–1.03)0.091
      P < 0.05 shown in bold.
      HR, hazard ratio.

      Multivariate Analysis of Parameters for Overall Survival

      The results of multivariate analysis of morphometric parameters and independent clinicopathologic parameters for overall survival are listed in Table 5. Among the 16 morphometric parameters with significant influence on overall survival in univariate analysis, only low skewness of nuclear irregularity (P = 0.0054) and high mean of nuclear irregularity (P = 0.033) remained significant adverse prognostic factors. Among the seven independent clinicopathologic parameters with significant adverse prognostic influence in univariate analysis, only age >60 years at diagnosis remained significant (P = 0.025). Of note, blastoid/pleomorphic morphology and Ki-67 ≥ 30% were no more significant in multivariate analysis.
      Table 5Influence of Morphometric Parameters (or the Morphometric Score) and Independent Clinicopathologic Parameters on Overall Survival by Multivariate Analysis
      ParameterHazard ratio (95% CI)P value
      Morphometric parameters and independent clinicopathologic parameters
       Low skewness of nuclear irregularity12.7 (2.13–76.5)0.0054
       High mean of nuclear irregularity5.90 (1.16–30.1)0.033
       High mean of nuclear width0.056
       High variance of nuclear length0.058
       High variance of nuclear perimeter0.15
       High mean of nuclear length0.20
       Low skewness of nuclear perimeter0.37
       Low skewness of nuclear length0.53
       Low kurtosis of nuclear length/width ratio0.55
       High mean of nuclear length/width ratio0.69
       High variance of nuclear area0.86
       Low kurtosis of nuclear irregularity0.88
       High mean of nuclear perimeter0.88
       Low kurtosis of nuclear perimeter0.88
       High variance of nuclear width0.90
       High mean of nuclear area0.91
       Age >60 years3.14 (1.15–8.54)0.025
       Elevated serum LDH2.02 (0.95–4.30)0.070
       Stage III or IV2.64 (0.76–9.21)0.13
       B symptoms1.81 (0.77–4.27)0.17
       ECOG score >11.78 (0.76–4.19)0.18
       Blastoid/pleomorphic morphology1.44 (0.55–3.77)0.46
       Ki-67 ≥ 30%0.88 (0.36–2.18)0.78
      Morphometric score and independent clinicopathologic parameters
       Morphometric score2.40 (1.35–4.25)0.0028
       Age >60 years1.69 (0.81–3.54)0.17
       Stage III or IV2.31 (0.80–6.64)0.12
       B symptoms2.07 (1.01–4.22)0.047
       Elevated serum LDH1.45 (0.77–2.73)0.26
       ECOG score >11.77 (0.82–3.83)0.15
       Blastoid/pleomorphic morphology1.05 (0.47–2.35)0.90
       Ki-67 ≥ 30%1.12 (0.54–2.34)0.75
      P < 0.05 shown in bold.
      ECOG, Eastern Cooperative Oncology Group; LDH, lactate dehydrogenase.
      The results of multivariate analysis of morphometric parameters, bMIPI, and non-bMIPI clinicopathologic parameters for overall survival are listed in Table 6. Only low skewness of nuclear irregularity [P = 0.020; hazard ratio (HR), 9.22; 95% CI, 1.42–59.7], high mean of nuclear irregularity (P = 0.047; HR, 5.11; 95% CI, 1.02–25.5), and bMIPI risk group (P = 0.025; HR, 1.80; 95% CI, 1.08–3.01) remained significant adverse prognostic factors. All other morphometric and clinicopathologic parameters, including blastoid/pleomorphic morphology, had no more significant influence on survival in multivariate analysis.
      Table 6Influence of Morphometric Parameters (or the Morphometric Score), bMIPI, and Non-bMIPI Clinicopathologic Parameters on Overall Survival by Multivariate Analysis
      ParameterHazard ratio (95% CI)P value
      Morphometric parameters, bMIPI, and non-bMIPI clinicopathologic parameters
       Low skewness of nuclear irregularity9.22 (1.42–59.7)0.020
       High mean of nuclear irregularity5.11 (1.02–25.5)0.047
       High mean of nuclear length0.15
       High variance of nuclear length0.17
       High variance of nuclear perimeter0.19
       High mean of nuclear width0.28
       Low kurtosis of nuclear perimeter0.41
       Low kurtosis of nuclear irregularity0.60
       High mean of nuclear length/width ratio0.63
       Low kurtosis of nuclear length/width ratio0.78
       Low skewness of nuclear length0.83
       High variance of nuclear area0.86
       High mean of nuclear perimeter0.88
       High variance of nuclear width0.89
       High mean of nuclear area0.90
       Low skewness of nuclear perimeter0.91
       bMIPI risk group1.80 (1.08–3.01)0.025
       Stage III or IV3.14 (0.93–10.7)0.066
       B symptoms1.56 (0.71–3.43)0.27
       Blastoid/pleomorphic morphology1.25 (0.56–2.79)0.59
      Morphometric score and independent clinicopathologic parameters
       Morphometric score2.32 (1.31–4.11)0.0038
       bMIPI risk group1.66 (1.06–2.58)0.025
       Stage III or IV2.24 (0.81–6.17)0.12
       B symptoms1.97 (1.00–3.87)0.050
       Blastoid/pleomorphic morphology1.07 (0.54–2.12)0.85
      P < 0.05 shown in bold.
      bMIPI, Biologic Mantle Cell Lymphoma International Prognostic Index.

      Morphometric Score and Overall Survival

      We calculated a morphometric score using the two morphometric parameters with independent adverse prognostic influence in multivariate analysis. One point was assigned for each of the following risk factors:low skewness of nuclear irregularity (≤3.729) andhigh mean of nuclear irregularity (>0.1304 μm2).
      The morphometric score, ranging from 0 to 2, was an adverse prognostic factor in univariate analysis for overall survival (P = 0.0011; HR, 2.34; 95% CI, 1.40–3.90). The survival curves of patients stratified by the morphometric score are shown in Figure 2J. Multivariate analysis of the morphometric score and independent clinicopathologic parameters showed that only the morphometric score (P = 0.0028) and B symptoms (P = 0.047) were independent adverse prognostic factor (Table 5). Multivariate analysis of the morphometric score, bMIPI, and non-bMIPI clinicopathologic parameters showed that only the morphometric score (P = 0.0038; HR, 2.32; 95% CI, 1.31–4.11) and bMIPI risk group (P = 0.025; HR, 1.66; 95% CI, 1.06–2.58) remained independent adverse prognostic factors (Table 6).

      Visualization of Morphometric Score

      Density plots were produced to demonstrate the distribution of nuclear irregularity in each case. Examples of classic MCL, blastoid MCL, and pleomorphic MCL with different morphometric scores are shown in Figure 4, A–C, respectively. Higher mean of nuclear irregularity and lower skewness of nuclear irregularity result in a higher morphometric score. With increase of the morphometric score, there is a trend of increasingly pleomorphic nuclei in the microscopic images. However, finding an optimal cutoff point of nuclear pleomorphism with the highest prognostic significance by human eyes is difficult. The density plots of nuclear irregularity distribution and microscopic images of all cases can be found in Supplemental Figure S1.
      Figure thumbnail gr4
      Figure 4Examples of cases with different morphometric scores. AC: The microscopic images with tumor cell nuclei (yellow closed lines) automatically delineated by our model and density plots of nuclear irregularity (median = blue dotted vertical line) of cases with classic (A), blastoid (B), and pleomorphic (C) morphology. A higher mean (red vertical line) and a lower skewness (less concentration of data at the left in the density plot) resulted in a higher morphometric score. The nuclei in the microscopic images appeared increasingly pleomorphic with increase of the morphometric score. D: All cases of blastoid and pleomorphic mantle cell lymphoma (MCL) had a high morphometric score of 2, whereas classic MCL cases had a morphometric score of 0, 1, or 2.

      Correlation of Morphometric Score with Clinicopathologic Features

      The comparison of clinicopathologic features between cases with a low (0 or 1) and high (2) morphometric score is listed in Table 7. A high morphometric score was significantly associated with blastoid/pleomorphic morphology (P = 0.0012), Ki-67 ≥ 30% (P = 0.0025), and high bMIPI risk group (P = 0.0079). Of note, all 26 cases with blastoid or pleomorphic morphology had a high morphometric score of 2 (Figure 4D).
      Table 7Comparison of Clinicopathologic Features between Cases With a Low and a High MS
      ParameterLow MS (MS = 0 or 1; n = 24)High MS (MS = 2; n = 79)P value
      Age at diagnosis, years
       ≤6013 (54.2)26 (32.9)
       >6011 (45.8)53 (67.1)0.060
      Sex
       Female5 (20.8)12 (15.2)
       Male19 (79.2)67 (84.8)0.74
      Stage
      Some cases excluded because of incomplete clinical data.
       I or II3 (13.0)8 (10.7)
       III or IV20 (87.0)67 (89.3)1.0
      B symptoms
      Some cases excluded because of incomplete clinical data.
       No16 (80.0)50 (75.8)
       Yes4 (20.0)16 (24.2)0.93
      Serum LDH
      Some cases excluded because of incomplete clinical data.
       Normal15 (65.2)37 (52.1)
       Elevated8 (34.8)34 (47.9)0.27
      ECOG score
      Some cases excluded because of incomplete clinical data.
       0 or 121 (91.3)52 (77.6)
       >12 (8.7)15 (22.4)0.26
      Extranodal site
      Some cases excluded because of incomplete clinical data.
       0 or 18 (34.8)26 (37.7)
       >115 (65.2)43 (62.3)0.80
      BM/PB involvement
      Some cases excluded because of incomplete clinical data.
       No7 (30.4)26 (40.6)
       Yes16 (69.6)38 (59.4)0.39
      MIPI risk group
      Some cases excluded because of incomplete clinical data.
       Low to intermediate14 (63.6)33 (50.8)
       High8 (36.4)32 (49.2)0.30
      Morphologic variant
       Classic24 (100)53 (67.1)
       Blastoid/pleomorphic0 (0)26 (32.9)0.0012
      Cyclin D1
       Negative0 (0)6 (7.6)
       Positive24 (100)73 (92.4)0.37
      Ki-67, %
       <3021 (87.5)42 (53.2)
       ≥303 (12.5)37 (46.8)0.0025
      bMIPI risk group
      Some cases excluded because of incomplete clinical data.
       Low to intermediate16 (72.7)26 (40.0)
       High6 (27.3)39 (60.0)0.0079
      P < 0.05 shown in bold. Data are given as number (percentage).
      bMIPI, Biologic MIPI; BM/PB, bone marrow/peripheral blood; ECOG, Eastern Cooperative Oncology Group; LDH, lactate dehydrogenase; MIPI, Mantle Cell Lymphoma International Prognostic Index; MS, morphometric score.
      Some cases excluded because of incomplete clinical data.

      Morphometric bMIPI Score and Overall Survival

      Because the morphometric score and bMIPI risk group were independent adverse prognostic factors in multivariate analysis, we calculated a morphometric bMIPI score by adding the morphometric score with a bMIPI score (low risk, 0; intermediate risk, 1; and high risk, 2). The morphometric bMIPI score, ranging from 0 to 4, was useful for risk stratification of patients with MCL, with a P value of 0.000009 (Figure 2K). Alternatively, a three-tiered risk grouping (low, 0 to 1; intermediate, 2 to 3; and high, 4) achieved a P value of 0.000001 (Figure 2L).

      Discussion

      Blastoid/pleomorphic morphology has long been known to correlate with worse prognosis in patients with MCL.
      • Bernard M.
      • Gressin R.
      • Lefrere F.
      • Drenou B.
      • Branger B.
      • Caulet-Maugendre S.
      • Tass P.
      • Brousse N.
      • Valensi F.
      • Milpied N.
      • Voilat L.
      • Sadoun A.
      • Ghandour C.
      • Hunault M.
      • Leloup R.
      • Mannone L.
      • Hermine O.
      • Lamy T.
      Blastic variant of mantle cell lymphoma: a rare but highly aggressive subtype.
      However, identification of these morphologic variants could be challenging for general pathologists with less experience.
      • Dreyling M.
      • Klapper W.
      • Rule S.
      Blastoid and pleomorphic mantle cell lymphoma: still a diagnostic and therapeutic challenge!.
      In a recent study, there was only a moderate concordance (κ = 0.57) in identification of blastoid/pleomorphic morphology in MCL among eight reference pathology laboratories of the European MCL Network,
      • Croci G.A.
      • Hoster E.
      • Bea S.
      • Clot G.
      • Enjuanes A.
      • Scott D.W.
      • Cabecadas J.
      • Veloza L.
      • Campo E.
      • Clasen-Linde E.
      • Goswami R.S.
      • Helgeland L.
      • Pileri S.
      • Rymkiewicz G.
      • Reinke S.
      • Dreyling M.
      • Klapper W.
      Reproducibility of histologic prognostic parameters for mantle cell lymphoma: cytology, Ki67, p53 and SOX11.
      indicating the difficulty even for expert hematopathologists.
      Theoretically, similar to our previous studies,
      • Chuang W.Y.
      • Chang S.H.
      • Yu W.H.
      • Yang C.K.
      • Yeh C.J.
      • Ueng S.H.
      • Liu Y.J.
      • Chen T.D.
      • Chen K.H.
      • Hsieh Y.Y.
      • Hsia Y.
      • Wang T.H.
      • Hsueh C.
      • Kuo C.F.
      • Yeh C.Y.
      Successful identification of nasopharyngeal carcinoma in nasopharyngeal biopsies using deep learning.
      ,
      • Chuang W.Y.
      • Chen C.C.
      • Yu W.H.
      • Yeh C.J.
      • Chang S.H.
      • Ueng S.H.
      • Wang T.H.
      • Hsueh C.
      • Kuo C.F.
      • Yeh C.Y.
      Identification of nodal micrometastasis in colorectal cancer using deep learning on annotation-free whole-slide images.
      a direct end-to-end training process could be used to establish a deep learning model to identify aggressive morphologic variants of MCL. However, because MCL is a rare type of B-cell lymphoma
      • Swerdlow S.H.
      • Campo E.
      • Seto M.
      • Müller-Hermelink H.K.
      and its blastoid and pleomorphic variants are even rarer, it is difficult to collect numerous cases of each variant for such end-to-end training. Recently, it has been shown that morphometric parameters can be used to classify primary intestinal T-cell lymphoma.
      • Yu W.H.
      • Li C.H.
      • Wang R.C.
      • Yeh C.Y.
      • Chuang S.S.
      Machine learning based on morphological features enables classification of primary intestinal T-cell lymphomas.
      Herein, we analyzed the disease prognosis based on objective morphometric parameters extracted from tens of thousands of nuclei among 103 patients with MCL to derive a significant and human-interpretable result.
      Morphometric studies require accurate contouring of targets. High-quality segmentation of nuclei is essential for correct measurement of nuclear attributes. Previously, nuclear segmentation was performed using classic image segmentation methods, such as thresholding, region-based approaches, energy minimization techniques, and classification-based segmentation.
      • Vahadane A.
      • Sethi A.
      Towards generalized nuclear segmentation in histological images. 13th IEEE International Conference on BioInformatics and BioEngineering (BIBE).
      However, these traditional methods often assume a certain image pattern of nuclei and are prone to failure when the assumption is violated. Recently, deep learning has been successful in detection,
      • Chuang W.Y.
      • Chang S.H.
      • Yu W.H.
      • Yang C.K.
      • Yeh C.J.
      • Ueng S.H.
      • Liu Y.J.
      • Chen T.D.
      • Chen K.H.
      • Hsieh Y.Y.
      • Hsia Y.
      • Wang T.H.
      • Hsueh C.
      • Kuo C.F.
      • Yeh C.Y.
      Successful identification of nasopharyngeal carcinoma in nasopharyngeal biopsies using deep learning.
      ,
      • Chuang W.Y.
      • Chen C.C.
      • Yu W.H.
      • Yeh C.J.
      • Chang S.H.
      • Ueng S.H.
      • Wang T.H.
      • Hsueh C.
      • Kuo C.F.
      • Yeh C.Y.
      Identification of nodal micrometastasis in colorectal cancer using deep learning on annotation-free whole-slide images.
      classification,
      • Chen C.L.
      • Chen C.C.
      • Yu W.H.
      • Chen S.H.
      • Chang Y.C.
      • Hsu T.I.
      • Hsiao M.
      • Yeh C.Y.
      • Chen C.Y.
      An annotation-free whole-slide training approach to pathological classification of lung cancer types using deep learning.
      ,
      • Steinbuss G.
      • Kriegsmann M.
      • Zgorzelski C.
      • Brobeil A.
      • Goeppert B.
      • Dietrich S.
      • Mechtersheimer G.
      • Kriegsmann K.
      Deep learning for the classification of non-hodgkin lymphoma on histopathological images.
      and grading
      • Bulten W.
      • Pinckaers H.
      • van Boven H.
      • Vink R.
      • de Bel T.
      • van Ginneken B.
      • van der Laak J.
      • Hulsbergen-van de Kaa C.
      • Litjens G.
      Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study.
      ,
      • Strom P.
      • Kartasalo K.
      • Olsson H.
      • Solorzano L.
      • Delahunt B.
      • Berney D.M.
      • Bostwick D.G.
      • Evans A.J.
      • Grignon D.J.
      • Humphrey P.A.
      • Iczkowski K.A.
      • Kench J.G.
      • Kristiansen G.
      • van der Kwast T.H.
      • Leite K.R.M.
      • McKenney J.K.
      • Oxley J.
      • Pan C.C.
      • Samaratunga H.
      • Srigley J.R.
      • Takahashi H.
      • Tsuzuki T.
      • Varma M.
      • Zhou M.
      • Lindberg J.
      • Lindskog C.
      • Ruusuvuori P.
      • Wahlby C.
      • Gronberg H.
      • Rantalainen M.
      • Egevad L.
      • Eklund M.
      Artificial intelligence for diagnosis and grading of prostate cancer in biopsies: a population-based, diagnostic study.
      of tumors in pathology images. Nowadays, the best performing nuclear segmentation methods are all deep learning based, and their performance is comparable to that of human annotators.
      • Kumar N.
      • Verma R.
      • Anand D.
      • Zhou Y.
      • Onder O.F.
      • Tsougenis E.
      • et al.
      A multi-organ nucleus segmentation challenge.
      Indeed, our deep learning–based nuclear segmentation model achieved high performance, thus enabling accurate morphometry (Figure 3 and Supplemental Figure S1). The average aggregated Jaccard index of our predicted nuclei ranged from 0.685 to 0.696 in different morphologic subtypes, similar to that of a previously reported best preforming deep learning–based nuclear segmentation model (average aggregated Jaccard index = 0.691).
      • Kumar N.
      • Verma R.
      • Anand D.
      • Zhou Y.
      • Onder O.F.
      • Tsougenis E.
      • et al.
      A multi-organ nucleus segmentation challenge.
      More than three decades ago, morphometric studies demonstrated that lymphoid cells in different areas of normal lymphoid tissue had different nuclear parameters.
      • Dardick I.
      Morphometry of normal human lymphoid tissues: nuclear invaginations and clefts.
      ,
      • Dardick I.
      • Dardick A.M.
      Morphometry of normal human lymphoid tissues: nuclear parameters for comparative studies of lymphoma.
      Other old studies investigated the nuclear morphometry of non-Hodgkin lymphoma and its potential use in subclassfication.
      • Crocker J.
      • Jones E.L.
      • Curran R.C.
      A comparative study of nuclear form factor, area and diameter in non-Hodgkin's lymphomas and reactive lymph nodes.
      • Gupta S.
      • Gupta R.
      • Singh S.
      • Gupta K.
      • Kudesia M.
      Nuclear morphometry and texture analysis of B-cell non-Hodgkin lymphoma: utility in subclassification on cytosmears.
      • Roy S.
      • Joshi K.
      • Choudhury A.R.
      • Varma S.C.
      • Banerjee C.K.
      Nuclear morphometric analysis of non-Hodgkin's lymphoma.
      Later on, a study showed that the nuclear area of MCL tumor cells was significantly larger than that of mantle zone lymphocytes in reactive tonsils.
      • Baek T.
      • Huh J.
      • Kwak H.
      • Park M.
      • Lee H.
      Image analytic study of nuclear area in mantle cell lymphoma.
      No cases with blastoid/pleomorphic morphology were included in that study, and their mean nuclear area of MCL measured with manual annotation by three pathologists was 37.9, 37.9, and 38.2 μm2, respectively.
      • Baek T.
      • Huh J.
      • Kwak H.
      • Park M.
      • Lee H.
      Image analytic study of nuclear area in mantle cell lymphoma.
      Their result was similar to that of our classic MCL (42.42 μm2). However, no other morphometric parameters were analyzed in that study, and correlation with survival was not investigated.
      • Baek T.
      • Huh J.
      • Kwak H.
      • Park M.
      • Lee H.
      Image analytic study of nuclear area in mantle cell lymphoma.
      Compared with classic MCL, we found that 16 and 13 morphometric parameters related to nuclear size and shape were significantly different in blastoid and pleomorphic MCL, respectively (Table 3). Both aggressive variants had higher mean, higher variance, lower skewness, and/or lower kurtosis of nuclear length, width, perimeter, area, and/or irregularity. Higher mean and variance of these morphometric attributes indicate larger and more irregular nuclei with more variation in size and shape. Lower skewness reflects less concentration of data at the left in the density plot, whereas lower kurtosis correlates with more data in shoulders other than the peak or tails in distribution.
      • Westfall P.H.
      Kurtosis as peakedness, 1905 - 2014. R.I.P.
      Both lower skewness and lower kurtosis of morphometric attributes also contribute to the increased pleomorphism perceived by human eyes.
      Theoretically, the more dispersed chromatin in blastoid or pleomorphic MCL should correlate with higher pixel randomness (namely, higher entropy). However, no significant difference in entropy was found between these aggressive variants and classic MCL (Table 3). Blastoid and pleomorphic variants are known to have more prominent nucleoli, which result in local aggregation of pixels and decrease of pixel randomness. This could explain the lack of significant difference in nuclear entropy between classic MCL and blastoid or pleomorphic MCL.
      Using univariate analysis, 16 of the 32 morphometric parameters related to nuclear size and shape had significant influence on survival (Table 4). Of note, 14 (87.5%) of them were significantly different between classic MCL and blastoid or pleomorphic MCL (Table 3). Similar to the difference observed in blastoid and pleomorphic variants of MCL compared with classic MCL, higher mean, higher variance, lower skewness, and lower kurtosis of these morphometric attributes were associated with poor prognosis. Our results suggest that morphometry could objectively quantify pleomorphism with prognostic significance.
      To better evaluate the robustness of the results mentioned above, a bootstrapping (with 1000 repetitions) procedure was performed for prognostic analysis of each morphometric parameter. All 16 morphometric parameters remained prognostically significant after our bootstrap simulation, demonstrating the robustness of their prognostic effect.
      Using multivariate analysis, we selected two morphometric parameters with independent prognostic value (mean and skewness of nuclear irregularity) (Tables 5 and 6) to calculate a morphometric score. Although a high morphometric score was strongly associated with blastoid/pleomorphic morphology (P = 0.0012) (Table 7), the morphometric score but not blastoid/pleomorphic morphology had independent prognostic value in multivariate analysis (Tables 5 and 6). In fact, a high morphometric score of 2 identified not only all 26 blastoid/pleomorphic cases but also 53 cases (68.8%) of classic MCL with worse prognosis (Figures 2J and 4D). In addition, a low morphometric score of 0 identified 9 cases (11.7%) of classic MCL with good prognosis (Figures 2J and 4D). Our morphometric method not only objectively quantified the pleomorphism of tumor cell nuclei but also revealed a better cutoff value of pleomorphism with higher prognostic significance than the blastoid/pleomorphic morphology identified by human eyes.
      Interestingly, a previous morphometric study on follicular lymphoma showed higher mean, higher SD (square root of variance), lower skewness, and lower kurtosis of nuclear area in cases composed of large cells compared with those composed of small cleaved cells.
      • Stevens M.W.
      • Crowley K.S.
      • Fazzalari N.L.
      • Woods A.E.
      Use of morphometry in cytological preparations for diagnosing follicular non-Hodgkin's lymphomas.
      Their finding suggests a correlation between these morphometric differences and high tumor grade in follicular lymphoma. It bears some resemblance to our finding that high mean, high variance, low skewness, and low kurtosis of nuclear size and irregularity are associated with poor prognosis in MCL. Future study is needed to clarify whether such phenomenon exists in other types of malignant lymphoma.
      Although blastoid/pleomorphic morphology is associated with poor prognosis in MCL, it has been shown that its prognostic significance is overruled by the proliferation index Ki-67 in multivariate analysis.
      • Hoster E.
      • Rosenwald A.
      • Berger F.
      • Bernd H.W.
      • Hartmann S.
      • Loddenkemper C.
      • Barth T.F.
      • Brousse N.
      • Pileri S.
      • Rymkiewicz G.
      • Kodet R.
      • Stilgenbauer S.
      • Forstpointner R.
      • Thieblemont C.
      • Hallek M.
      • Coiffier B.
      • Vehling-Kaiser U.
      • Bouabdallah R.
      • Kanz L.
      • Pfreundschuh M.
      • Schmidt C.
      • Ribrag V.
      • Hiddemann W.
      • Unterhalt M.
      • Kluin-Nelemans J.C.
      • Hermine O.
      • Dreyling M.H.
      • Klapper W.
      Prognostic value of Ki-67 index, cytology, and growth pattern in mantle-cell lymphoma: results from randomized trials of the european mantle cell lymphoma network.
      Unlike blastoid/pleomorphic morphology, which is identified by human eyes, our morphometric parameters can be objectively measured. In addition, the prognostic value of our morphometric score overrode blastoid/pleomorphic morphology in multivariate analysis. Combining the morphometric score and bMIPI risk group, our morphometric bMIPI score performed well in risk stratification of patients with MCL (P = 0.000001), better than bMIPI (P = 0.00016) and MIPI (P = 0.00099) (Figure 2).
      In conclusion, we discovered a new independent prognostic factor in MCL using deep learning–based nuclear morphometry. Such a morphometric score can be objectively measured, and its prognostic value is more robust than that of blastoid/pleomorphic morphology. Future large-scale prospective studies would be helpful to confirm our results.

      Supplemental Data

      • Supplemental Figure S1

        Representative regions of interest and density plots of nuclear irregularity (mean, red line; median, blue dotted line) of all cases sorted by morphometric score. The nuclear contours (yellow closed lines) of tumor cells were automatically delineated by our model.

      • Supplementary Figure S2

        Left panels: The distribution of each nuclear morphometric parameter with a cutoff value (vertical line) determined by the software valuate Cutpoints.

        • Ogluszka M.
        • Orzechowska M.
        • Jedroszka D.
        • Witas P.
        • Bednarek A.K.
        Evaluate cutpoints: adaptable continuous data distribution system for determining survival in Kaplan-Meier estimator.
        Right panels: The survival curves stratified by each morphometric parameter using the cutoff value are also shown. P < 0.05 is shown in red.

      References

        • Swerdlow S.H.
        • Campo E.
        • Seto M.
        • Müller-Hermelink H.K.
        Swerdlow S.H. Campo E. Harris N.L. Jaffe E.S. Pileri S.A. Stein H. Thiele J. Mantle cell lymphoma. WHO Classification of the Tumours of Haematopoietic and Lymphoid Tissues. Revised ed 4. IARC, Lyon, France2017: 285-290
        • Fu K.
        • Weisenburger D.D.
        • Greiner T.C.
        • Dave S.
        • Wright G.
        • Rosenwald A.
        • Chiorazzi M.
        • Iqbal J.
        • Gesk S.
        • Siebert R.
        • De Jong D.
        • Jaffe E.S.
        • Wilson W.H.
        • Delabie J.
        • Ott G.
        • Dave B.J.
        • Sanger W.G.
        • Smith L.M.
        • Rimsza L.
        • Braziel R.M.
        • Muller-Hermelink H.K.
        • Campo E.
        • Gascoyne R.D.
        • Staudt L.M.
        • Chan W.C.
        • Lymphoma/Leukemia Molecular Profiling Project
        Cyclin D1-negative mantle cell lymphoma: a clinicopathologic study based on gene expression profiling.
        Blood. 2005; 106: 4315-4321
        • Martin-Garcia D.
        • Navarro A.
        • Valdes-Mas R.
        • Clot G.
        • Gutierrez-Abril J.
        • Prieto M.
        • et al.
        CCND2 and CCND3 hijack immunoglobulin light-chain enhancers in cyclin D1(-) mantle cell lymphoma.
        Blood. 2019; 133: 940-951
        • Chuang W.Y.
        • Chang H.
        • Chang G.J.
        • Wang T.H.
        • Chang Y.S.
        • Wang T.H.
        • Yeh C.J.
        • Ueng S.H.
        • Chien H.P.
        • Chang C.Y.
        • Wan Y.L.
        • Hsueh C.
        Pleomorphic mantle cell lymphoma morphologically mimicking diffuse large B cell lymphoma: common cyclin D1 negativity and a simple immunohistochemical algorithm to avoid the diagnostic pitfall.
        Histopathology. 2017; 70: 986-999
        • Chuang W.Y.
        • Chang S.T.
        • Yuan C.T.
        • Chang G.J.
        • Chang H.
        • Yeh C.J.
        • Ueng S.H.
        • Kao H.W.
        • Wang T.H.
        • Wan Y.L.
        • Shih L.Y.
        • Chuang S.S.
        • Hsueh C.
        Identification of CD5/cyclin D1 double-negative pleomorphic mantle cell lymphoma: a clinicopathologic, genetic, and gene expression study.
        Am J Surg Pathol. 2020; 44: 232-240
        • Mozos A.
        • Royo C.
        • Hartmann E.
        • De Jong D.
        • Baro C.
        • Valera A.
        • Fu K.
        • Weisenburger D.D.
        • Delabie J.
        • Chuang S.S.
        • Jaffe E.S.
        • Ruiz-Marcellan C.
        • Dave S.
        • Rimsza L.
        • Braziel R.
        • Gascoyne R.D.
        • Sole F.
        • Lopez-Guillermo A.
        • Colomer D.
        • Staudt L.M.
        • Rosenwald A.
        • Ott G.
        • Jares P.
        • Campo E.
        SOX11 expression is highly specific for mantle cell lymphoma and identifies the cyclin D1-negative subtype.
        Haematologica. 2009; 94: 1555-1562
        • Zeng W.
        • Fu K.
        • Quintanilla-Fend L.
        • Lim M.
        • Ondrejka S.
        • Hsi E.D.
        Cyclin D1-negative blastoid mantle cell lymphoma identified by SOX11 expression.
        Am J Surg Pathol. 2012; 36: 214-219
        • Hoster E.
        • Dreyling M.
        • Klapper W.
        • Gisselbrecht C.
        • van Hoof A.
        • Kluin-Nelemans H.C.
        • Pfreundschuh M.
        • Reiser M.
        • Metzner B.
        • Einsele H.
        • Peter N.
        • Jung W.
        • Wormann B.
        • Ludwig W.D.
        • Duhrsen U.
        • Eimermacher H.
        • Wandt H.
        • Hasford J.
        • Hiddemann W.
        • Unterhalt M.
        • German Low Grade Lymphoma Study Group (GLSG), European Mantle Cell Lymphoma Network
        A new prognostic index (MIPI) for patients with advanced-stage mantle cell lymphoma.
        Blood. 2008; 111: 558-565
        • Royo C.
        • Navarro A.
        • Clot G.
        • Salaverria I.
        • Gine E.
        • Jares P.
        • Colomer D.
        • Wiestner A.
        • Wilson W.H.
        • Vegliante M.C.
        • Fernandez V.
        • Hartmann E.M.
        • Trim N.
        • Erber W.N.
        • Swerdlow S.H.
        • Klapper W.
        • Dyer M.J.
        • Vargas-Pabon M.
        • Ott G.
        • Rosenwald A.
        • Siebert R.
        • Lopez-Guillermo A.
        • Campo E.
        • Bea S.
        Non-nodal type of mantle cell lymphoma is a specific biological and clinical subgroup of the disease.
        Leukemia. 2012; 26: 1895-1898
        • Hoster E.
        • Rosenwald A.
        • Berger F.
        • Bernd H.W.
        • Hartmann S.
        • Loddenkemper C.
        • Barth T.F.
        • Brousse N.
        • Pileri S.
        • Rymkiewicz G.
        • Kodet R.
        • Stilgenbauer S.
        • Forstpointner R.
        • Thieblemont C.
        • Hallek M.
        • Coiffier B.
        • Vehling-Kaiser U.
        • Bouabdallah R.
        • Kanz L.
        • Pfreundschuh M.
        • Schmidt C.
        • Ribrag V.
        • Hiddemann W.
        • Unterhalt M.
        • Kluin-Nelemans J.C.
        • Hermine O.
        • Dreyling M.H.
        • Klapper W.
        Prognostic value of Ki-67 index, cytology, and growth pattern in mantle-cell lymphoma: results from randomized trials of the european mantle cell lymphoma network.
        J Clin Oncol. 2016; 34: 1386-1394
        • Cheson B.D.
        • Fisher R.I.
        • Barrington S.F.
        • Cavalli F.
        • Schwartz L.H.
        • Zucca E.
        • Lister T.A.
        • Alliance A.L.
        • Lymphoma G, Eastern Cooperative Oncology G, European Mantle Cell Lymphoma C, Italian Lymphoma F, European Organisation for R, Treatment of Cancer/Dutch Hemato-Oncology G, Grupo Espanol de Medula O, German High-Grade Lymphoma Study G, German Hodgkin's Study G, Japanese Lymphorra Study G, Lymphoma Study A, Group NCT, Nordic Lymphoma Study G, Southwest Oncology G, United Kingdom National Cancer Research I
        Recommendations for initial evaluation, staging, and response assessment of Hodgkin and non-Hodgkin lymphoma: the Lugano classification.
        J Clin Oncol. 2014; 32: 3059-3068
        • Chen K.
        • Wang J.
        • Pang J.
        • Cao Y.
        • Xiong Y.
        • Li X.
        • Sun S.
        MMDetection: open MMLab detection toolbox and benchmark.
        (arXiv 2019. [Preprint])
        • Lin T.Y.
        • Maire M.
        • Belongie S.
        • Bourdev L.
        • Girshick R.
        • Hays J.
        • Perona P.
        • Microsoft C.O.C.O.
        Common objects in context.
        (arXiv 2014. [Preprint])
        • Chen K.
        • Pang J.
        • Wang J.
        • Xiong Y.
        • Li X.
        • Sun S.
        • Feng W.
        • Liu Z.
        • Shi J.
        • Ouyang W.
        • Loy C.C.
        • Lin D.
        Hybrid task cascade for instance segmentation. 2019 IEEE/CVF Conference on computer vision and pattern recognition (CVPR).
        2019: 4969-4978
        • He K.
        • Zhang X.
        • Ren S.
        • Sun J.
        Deep residual learning for image recognition. 2016 IEEE Conference on computer vision and pattern recognition (CVPR).
        2016: 770-778
        • Shorten C.
        • Khoshgoftaar T.M.
        A survey on image data augmentation for deep learning.
        J Big Data. 2019; 6: 60
        • Ogluszka M.
        • Orzechowska M.
        • Jedroszka D.
        • Witas P.
        • Bednarek A.K.
        Evaluate cutpoints: adaptable continuous data distribution system for determining survival in Kaplan-Meier estimator.
        Comput Methods Programs Biomed. 2019; 177: 133-139
        • Bernard M.
        • Gressin R.
        • Lefrere F.
        • Drenou B.
        • Branger B.
        • Caulet-Maugendre S.
        • Tass P.
        • Brousse N.
        • Valensi F.
        • Milpied N.
        • Voilat L.
        • Sadoun A.
        • Ghandour C.
        • Hunault M.
        • Leloup R.
        • Mannone L.
        • Hermine O.
        • Lamy T.
        Blastic variant of mantle cell lymphoma: a rare but highly aggressive subtype.
        Leukemia. 2001; 15: 1785-1791
        • Dreyling M.
        • Klapper W.
        • Rule S.
        Blastoid and pleomorphic mantle cell lymphoma: still a diagnostic and therapeutic challenge!.
        Blood. 2018; 132: 2722-2729
        • Croci G.A.
        • Hoster E.
        • Bea S.
        • Clot G.
        • Enjuanes A.
        • Scott D.W.
        • Cabecadas J.
        • Veloza L.
        • Campo E.
        • Clasen-Linde E.
        • Goswami R.S.
        • Helgeland L.
        • Pileri S.
        • Rymkiewicz G.
        • Reinke S.
        • Dreyling M.
        • Klapper W.
        Reproducibility of histologic prognostic parameters for mantle cell lymphoma: cytology, Ki67, p53 and SOX11.
        Virchows Arch. 2020; 477: 259-267
        • Chuang W.Y.
        • Chang S.H.
        • Yu W.H.
        • Yang C.K.
        • Yeh C.J.
        • Ueng S.H.
        • Liu Y.J.
        • Chen T.D.
        • Chen K.H.
        • Hsieh Y.Y.
        • Hsia Y.
        • Wang T.H.
        • Hsueh C.
        • Kuo C.F.
        • Yeh C.Y.
        Successful identification of nasopharyngeal carcinoma in nasopharyngeal biopsies using deep learning.
        Cancers (Basel). 2020; 12: 507
        • Chuang W.Y.
        • Chen C.C.
        • Yu W.H.
        • Yeh C.J.
        • Chang S.H.
        • Ueng S.H.
        • Wang T.H.
        • Hsueh C.
        • Kuo C.F.
        • Yeh C.Y.
        Identification of nodal micrometastasis in colorectal cancer using deep learning on annotation-free whole-slide images.
        Mod Pathol. 2021; 34: 1901-1911
        • Yu W.H.
        • Li C.H.
        • Wang R.C.
        • Yeh C.Y.
        • Chuang S.S.
        Machine learning based on morphological features enables classification of primary intestinal T-cell lymphomas.
        Cancers. 2021; 13: 5463
        • Vahadane A.
        • Sethi A.
        Towards generalized nuclear segmentation in histological images. 13th IEEE International Conference on BioInformatics and BioEngineering (BIBE).
        2013: 1-4 (Chania, Greece)
        • Chen C.L.
        • Chen C.C.
        • Yu W.H.
        • Chen S.H.
        • Chang Y.C.
        • Hsu T.I.
        • Hsiao M.
        • Yeh C.Y.
        • Chen C.Y.
        An annotation-free whole-slide training approach to pathological classification of lung cancer types using deep learning.
        Nat Commun. 2021; 12: 1193
        • Steinbuss G.
        • Kriegsmann M.
        • Zgorzelski C.
        • Brobeil A.
        • Goeppert B.
        • Dietrich S.
        • Mechtersheimer G.
        • Kriegsmann K.
        Deep learning for the classification of non-hodgkin lymphoma on histopathological images.
        Cancers. 2021; 13: 2419
        • Bulten W.
        • Pinckaers H.
        • van Boven H.
        • Vink R.
        • de Bel T.
        • van Ginneken B.
        • van der Laak J.
        • Hulsbergen-van de Kaa C.
        • Litjens G.
        Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study.
        Lancet Oncol. 2020; 21: 233-241
        • Strom P.
        • Kartasalo K.
        • Olsson H.
        • Solorzano L.
        • Delahunt B.
        • Berney D.M.
        • Bostwick D.G.
        • Evans A.J.
        • Grignon D.J.
        • Humphrey P.A.
        • Iczkowski K.A.
        • Kench J.G.
        • Kristiansen G.
        • van der Kwast T.H.
        • Leite K.R.M.
        • McKenney J.K.
        • Oxley J.
        • Pan C.C.
        • Samaratunga H.
        • Srigley J.R.
        • Takahashi H.
        • Tsuzuki T.
        • Varma M.
        • Zhou M.
        • Lindberg J.
        • Lindskog C.
        • Ruusuvuori P.
        • Wahlby C.
        • Gronberg H.
        • Rantalainen M.
        • Egevad L.
        • Eklund M.
        Artificial intelligence for diagnosis and grading of prostate cancer in biopsies: a population-based, diagnostic study.
        Lancet Oncol. 2020; 21: 222-232
        • Kumar N.
        • Verma R.
        • Anand D.
        • Zhou Y.
        • Onder O.F.
        • Tsougenis E.
        • et al.
        A multi-organ nucleus segmentation challenge.
        IEEE Trans Med Imaging. 2020; 39: 1380-1391
        • Dardick I.
        Morphometry of normal human lymphoid tissues: nuclear invaginations and clefts.
        Arch Pathol Lab Med. 1984; 108: 197-201
        • Dardick I.
        • Dardick A.M.
        Morphometry of normal human lymphoid tissues: nuclear parameters for comparative studies of lymphoma.
        Arch Pathol Lab Med. 1984; 108: 190-196
        • Crocker J.
        • Jones E.L.
        • Curran R.C.
        A comparative study of nuclear form factor, area and diameter in non-Hodgkin's lymphomas and reactive lymph nodes.
        J Clin Pathol. 1983; 36: 298-302
        • Gupta S.
        • Gupta R.
        • Singh S.
        • Gupta K.
        • Kudesia M.
        Nuclear morphometry and texture analysis of B-cell non-Hodgkin lymphoma: utility in subclassification on cytosmears.
        Diagn Cytopathol. 2010; 38: 94-103
        • Roy S.
        • Joshi K.
        • Choudhury A.R.
        • Varma S.C.
        • Banerjee C.K.
        Nuclear morphometric analysis of non-Hodgkin's lymphoma.
        Pathology. 1989; 21: 100-104
        • Baek T.
        • Huh J.
        • Kwak H.
        • Park M.
        • Lee H.
        Image analytic study of nuclear area in mantle cell lymphoma.
        Korean J Hematol. 2010; 45: 193-196
        • Westfall P.H.
        Kurtosis as peakedness, 1905 - 2014. R.I.P.
        Am Stat. 2014; 68: 191-195
        • Stevens M.W.
        • Crowley K.S.
        • Fazzalari N.L.
        • Woods A.E.
        Use of morphometry in cytological preparations for diagnosing follicular non-Hodgkin's lymphomas.
        J Clin Pathol. 1988; 41: 370-377