If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Osteosarcomas (OSs) are aggressive bone tumors with many divergent histologic patterns. During pathology review, OSs are subtyped based on the predominant histologic pattern; however, tumors often demonstrate multiple patterns. This high tumor heterogeneity coupled with scarcity of samples compared with other tumor types render histology-based prognosis of OSs challenging. To combat lower case numbers in humans, dogs with spontaneous OSs have been suggested as a model species. Herein, we adversarially train a convolutional neural network to classify distinct histologic patterns of OS in humans using mostly canine OS data during training. We show that adversarial training improves domain adaption of a histologic subtype classifier from canines to humans, achieving an average multiclass F1 score of 0.77 (95% CI, 0.74–0.79) and 0.80 (95% CI, 0.78–0.81) when compared with the ground truth in canines and humans, respectively. Finally, we applied our trained model to characterize the histologic landscape of 306 canine OSs and uncovered distinct clusters with markedly different clinical responses to standard-of-care therapy.
Osteosarcoma (OS) is a rare but aggressive pediatric malignancy with approximately 800 cases reported annually in the United States.
Patients with metastatic or relapsed disease have dismal outcomes, with survival rates of <30% despite aggressive salvage regimens that typically include additional surgery, radiotherapy, and chemotherapy with agents such as ifosfamide, etoposide, cyclophosphamide, gemcitabine, and topotecan.
This assessment is based on review of tumor sections harvested after local tumor control via surgery. Despite this, there is a subset of patients with high necrosis that still develop metastatic disease after completion of frontline therapy. Hence, additional prognostic biomarkers are needed for accurate prognosis prediction. Because naturally occurring canine osteosarcoma has strong biological, molecular, and histologic similarities to human osteosarcoma and is at least 10 times more common than human osteosarcoma, it can serve as a powerful translational model for cancer biomarker investigation and drug development.
In contrast to humans, the clinical workflow in dogs does not allow for assessment of response to neoadjuvant therapy, but rather access to the entire tumor at the time of diagnosis via limb amputation. This allows a greater area of untreated tumor for analysis and correlation with outcomes of that specific patient.
Furthermore, in canine OS, beyond tumor stage (ie, de novo metastatic disease), there are no known consistent prognostic features either within the primary tumor histology or other patient factors, such as tumor location, alkaline phosphatase status, and age/sex/breed.
This yielded a well-annotated canine OS data set in which to examine osteosarcoma histology and explore the potential of artificial intelligence (AI)–derived biomarkers. Specifically, we investigate whether techniques in AI using adversarial learning could support the development of a histologic subtype classifier for osteosarcomas that adapts from dogs to humans and a prognostic signature in dogs based on digital pathology whole slide images.
Tumors were biopsied pre-amputation and diagnosed as osteosarcoma by anatomic pathologists at Comparative Oncology Trials Consortium (COTC) institutions (https://ccr.cancer.gov/comparative-oncology-program/consortium, last accessed May 13, 2022). At the time of surgical limb amputation, additional tumor tissue was collected by COTC investigators as a part of the standard-of-care portion of the trial schema. All tumors were collected before any treatment. Dogs were randomized to receive either standard of care or standard of care + adjuvant sirolimus (rapamycin) therapy. Statistical analysis of the primary clinical outcomes of the entire cohort of dogs found no differences in disease-free interval or survival between the two arms; thus, cases were included together in the analysis presented herein. In addition, we obtained 39 human osteosarcoma samples from an in-house pathology residency training cohort. Of these 39 samples, only 11 were utilized in our study for validation of domain-agnostic features. Tumor tissue was placed in 10% neutral-buffered formalin for 24 hours and then subjected to EDTA slow decalcification. Tissue was then sectioned and stained with hematoxylin and eosin, according to standard histopathologic practice. Three canine cases were excluded from this study as slides from these cases were not available. Slides from remaining 306 canine cases and 39 human cases were digitized using Hamamatsu S60 digital scanner (Hamamatsu Photonics, Hamamatsu, Japan) in ×40 magnification or 0.23 μm per pixel. No additional manual quality control of surgical tumor specimen size or percentage tumor tissue was completed before data collection. The methods were performed in accordance with relevant guidelines and regulations and approved by each participating COTC veterinary institution that enrolled canine patients onto the clinical trials from which the image data were derived.
Annotation and Preprocessing of Whole Slide Image Data
Pathologist annotations for 95 dog slides and 11 human slides were obtained in xml format using HALO (Albuquerque, NM). Each annotation file contained coordinates of roughly marked region boundaries for each histologic subtype within each slide. Because osteoblastic subtype is the most dominant subtype in osteosarcoma, the main tumor areas were marked and annotated as osteoblastic. Any regions within this area exhibiting divergent histology were annotated as necrotic: vessel rich (VR), chondroblastic, fibroblastic, or giant cell rich.
Any unmarked regions falling outside main tumor areas were classified as other and consisted primarily of nontumor tissue, osteoid formations, and, in some cases, slide preparation artifacts, such as folded tissue and slide debris.
Training deep learning models on whole slide image tiles extracted from multiple magnifications has proven to be effective in a weakly supervised learning setting where region-level annotations by pathologists are not available and histologic features of interest are open ended.
The smallest regions of interest annotated by the pathologist have an area of approximately 25,000 μm2 and are represented by at least one tile of size 256 × 256 at ×10 magnification. A larger tile size would have resulted in fewer training tiles per histologic subtype, which would further increase class imbalance and cause overfitting, whereas a smaller tile size would have obscured important architectural features that go beyond cellular morphology (eg, tumor cells surrounding blood vessels, which are a characteristic feature of telangiectatic osteosarcoma). Hence, to train our image classification model, each whole slide image was scanned at ×10 magnification level and broken down into 256 × 256 pixel tiles.
Tiles containing >85% of white space were filtered out. Each remaining tile was assigned a single label based on any overlapping pathologist annotations. If a tile contained one or more tumor lesions of divergent histology (ie, a region exceeding 15% of the tile area), the tile was assigned the histologic class of the most dominant lesion (ie, the divergent lesion covering the highest percentage area). Otherwise, the tile was assigned label osteoblastic. For example, if a tile had 35% of its area marked as fibroblastic, then the tile gets assigned the label fibroblastic. If a tile is dominated by nontumor tissue or hemorrhage, it was assigned the label other. All other tiles from unmarked slides were regarded as unlabeled.
For training, we randomly selected 80% of all labeled tiles from dogs (source domain) and additionally 2000 randomly selected labeled tiles from humans (target domain). Of the remaining 20% labeled tiles from dogs, half were randomly selected for validation and hyperparameter tuning, and the remaining half were held out for testing along with the remaining labeled human tiles that were not selected for training. For reproducibility, we fixed the random seed in our codes generating the train, validation, and test splits. The distribution of tiles by histologic subtype and train, validation, and test split is shown in Figure 1 and Supplemental Table S1.
Before feeding a tile as input to the classification model, each tile was rescaled to 224 × 224 pixels, and its per-channel pixel intensities (ranging from 0 to 1) were normalized to follow a standard normal distribution using the following per-channel mean intensity and SDs estimated from the dog training data: mean (r = 0.8938, G = 0.5708, B = 0.7944) and SD (r = 0.1163, G = 0.1528, B = 0.0885). Furthermore, to artificially augment the size of the training set, each tile from a minibatch during training was flipped on one side at random.
Domain Adversarial Training of a Histologic Subtype Classification Model for Osteosarcomas
Let be examples from a source domain and be examples from a target domain where the number of examples available is typically much less than the number of examples available from the source domain. To train a classification model that adapts from the source domain to target domain, we extend the algorithm of Ganin and Lempitsky
to the supervised setting. Specifically, let be the parameter of the feature extraction backbone (ie, the function that takes as input an example and maps it to a set of features), let be the parameter of the subtype classifier (ie, the function that receives input from the feature extractor and predicts class label ), and let be the parameter of the domain classifier (ie, the function that receives input from the feature extractor and predicts the domain label ). Furthermore, let:
The first term in Equation 1 represents the subtype classification error, whereas the second term in Equation 1 represents the domain classification error and the hyperparameter controls the trade-off between the two errors. The goal of a domain adaption algorithm is then to find the saddle point of :
The domain classifier tries to minimize the domain classification error (because of the term), and the subtype classifier tries to minimize the subtype classification error. To find the saddle point, the domain classifier is trained adversarially with the label classifier. Consequently, the parameters of the feature extractor at the saddle point minimize the subtype classification error (ie, the learned features are discriminative) while maximizing the domain classification error (ie, the learned features are domain invariant). Adversarial training is implemented in practice by simply adding a gradient reversal layer just before the domain classifier and performing standard stochastic gradient descent (Figure 1). The update rule for the parameters after incorporating the gradient reversal layer is given by Equations 4, 5, and 6:
The hyperparameter represents the learning rate. To obtain a head start during training, we initialize the parameters of the feature extraction portion of the resnet50 convolutional neural network ( to the values obtained from pretraining resnet50 on the ImageNet data set.
With the help of stochastic gradient descent, we then simultaneously train the histologic subtype classifier and domain classifier over several epochs using the same resnet50 backbone to find parameters that get us closest to the saddle point of . To aid in faster convergence, we decrease the learning rate hyperparameter over each epoch, following Ganin and Lempitsky
The training batch size was set to 256 (sampling 32 patches per whole slide image in each batch). As an early stopping criterion, model training was halted after 15 epochs as the gap between train error and validation error begins to widen after 15 epochs. Hence, model training was halted after 15 epochs. The parameters achieving the best performance on the validation data set over 15 epochs were saved and eventually used for making predictions on held-out test data. The resnet50 architecture and training algorithm were implemented in python using PyTorch on an in-house dedicated server using a single Nvidia RTX A6000 GPU with 48 GB of VRAM.
Spatial Probability Map Generation and Burden Estimation for Each Histologic Subtype
To generate spatial probability maps, each whole slide image was processed by the trained patch-level histologic subtype classifier from left to right in a sliding window manner with a window size of 256 × 256 pixels and an overlap of 64 pixels. The resulting probability maps generated were further down sampled to ×5 base magnification via local average pooling of tile probabilities. We eventually generate six spatial probability maps: one for each class (excluding the other class, representing normal/benign/hemorrhagic tissue). The resulting probability maps can then be converted to gray scale or color images and visualized as shown in Figures 2 and 3.
Having generated spatial probability maps for each histologic subtype, one can then estimate its absolute burden in each patient's tumor while accounting for variable number of slides scanned per case using the following approach:
represents the probability of region i,j being classified a particular subtype. The summation term represents the total area. The term N in the denominator represents the number of slides scanned per case. We choose to quantify the absolute burden of each subtype instead of relative burden because each tumor was scanned at the same base magnification, and we had access to multiple slides scanned for each tumor in our cohort, including slides with tissue artifacts, such as folded tissue and osteoid formations. See Supplemental Table S2 for the estimated absolute burden of each subtype for all 306 canine cases analyzed in this study.
Data Preprocessing for K-Means Clustering Analysis
Given the estimated burden of each histologic subtype in each dog sample, we first center and scale the data and then perform a principal component analysis. The projections of each sample along the first two principal components, which capture most of the variability in the data, are then used for K-means clustering.
Implementation Details of K-Means Clustering and Survival Analysis
To perform K-means clustering, we used the kmeans() utility function implemented in R stats package with the following options set: maximum iterations = 500, and nstart (number of random initializations of cluster centers) = 100. For performing Kaplan-Meier and Cox proportional hazards regression analysis of the clinical data, we used the survfit() and cph() utility functions from the R survival package. Results of these analyses were plotted using the ggsurvplot() and ggforest utility functions from R survminer and GGally packages.
The code to train a classification model using domain adversarial learning, trained model weights, and scripts to reproduce the downstream results are available (https://github.com/spatkar94/adversarialdogs.git, last accessed September 30, 2022).
Overview of Whole Slide Imaging Cohorts Analyzed in this Study and the Adversarial Learning Approach
To precisely characterize the morphologic heterogeneity of osteosarcomas, we systematically collected and scanned 600 hematoxylin and eosin–stained slides of treatment-naïve primary tumors from a diverse collection of 306 dogs enrolled in a two-armed National Cancer Institute COTC clinical trial.
The distribution of dogs analyzed in this study by geographic location and breed is summarized in Supplemental Tables S3 and S4. In addition, 39 de-identified hematoxylin and eosin slides of human osteosarcomas were collected to evaluate species-agnostic histologic features. A veterinary anatomic pathologist (J.B.) annotated 95 and 11 slides from canine and human samples, respectively, to identify regions of necrosis or tumor-specific histologic patterns,
including osteoblastic, chondroblastic, fibroblastic, giant cell–rich, and VR regions. Unannotated regions were classified as other.
We then trained a resnet50 convolutional neural network on whole slide image patches of osteosarcoma to classify them into different histologic subtypes, necrosis, or nontumor areas in both dogs (source domain) and humans (target domain). Figure 1, A and B, and Supplemental Table S1 depict the distribution of whole slide image patches corresponding to each class in training, validation, and test data sets generated for dogs and humans, respectively. Patches from both the dog and human training set were simultaneously fed to a resnet50 convolutional neural network trained using a domain adversarial approach (Figure 1C), which encourages neural networks to learn features that are important for the classification task of interest while at the same time less sensitive to domain-specific differences in the data.
This was achieved by simultaneously training two classifiers that share the same feature extraction backbone. One classifier aimed to classify whole slide image patches into one of the predefined classes, whereas the other classifier aimed to distinguish the domain of each patch (ie, whether the patch comes from a dog or human sample). During training, the weights of the shared feature extraction backbone are updated such that we arrive at an equilibrium that minimizes classification error while maximizing domain error. Patches from the validation set were used to monitor for any signs of overfitting of the classification model (see Materials and Methods for more details). In the evaluation phase, patches from the held-out test set were evaluated using the trained histologic subtype classifier.
Adversarial Learning Improves Domain Adaptation of the Histologic Subtype Classifier from Dogs to Humans
Having trained a patch-level histologic subtype classification model in a domain adversarial manner, we next evaluated the performance of the trained model on held-out test whole slide image patches in both dogs and humans. To evaluate the model's performance, we computed the per-class precision, recall, and F1 scores obtained by comparing the model-predicted class labels of each whole slide image patch in the test set with the ground-truth labels obtained from overlapping pathologist annotations (see Materials and Methods). On average, the model achieved an F1-score of 0.77 (95% CI, 0.74–0.79) in dogs, and an F1-score of 0.8 (95% CI, 0.79–0.81) in humans (Figure 4, A–D ). Overall, the histologic subtype classification model adapts from dogs (source domain) to humans (target domain) after seeing <5% of labeled examples from the target domain. The subtype that had low precision (20%) and low recall (23%) on the target domain is the chondroblastic subtype and was most often confused with the more dominant osteoblastic subtype.
To evaluate the effect of domain adversarial training on model generalizability from source domain (dogs) to target domain (humans), we performed three control experiments: i) train the image classification model on labeled data from the source domain only and evaluate on target domain (transfer learning), ii) train the image classification model on labeled data from target domain only and evaluate on target domain, and iii) train the image classification model on labeled data from both the source and target domain using standard supervised learning and evaluate on target domain. For each experiment, the classification model was trained starting from the same set of initialized weights and hyperparameters. Overall, we found that the domain adversarial learning approach achieved significantly lower test error per epoch compared with the other three controls when evaluated on the target domain (Figure 4E).
To visualize the predictions of the patch-level histologic subtype classification model on the whole slide image, we generated spatial probability maps depicting regions of high versus low probability for each histologic subtype based on application of the patch-level histologic subtype classification model over the whole slide image in a sliding window manner (see Materials and Methods for details). As a qualitative validation, Figures 2 and 3 depict pathologist-marked region boundaries within four dog and two human osteosarcoma surgical specimens covering each histologic subtype along with classifier-derived probability maps (one per histologic subtype) over the whole slide image.
Unsupervised Exploratory Analysis of Whole Slide Imaging Features Uncovers Distinct Populations of Dogs with Different Responses to Standard-of-Care Therapy
Having generated spatial probability maps of each subtype, we next estimate the absolute burden of each subtype in each canine sample and apply the K-means clustering algorithm to identify clusters of dogs with similar whole slide tumor histology (Supplemental Table S2) (see Materials and Methods). Figure 5A plots the average silhouette score of inferred clusters for different values of K.
The higher the average silhouette score, the more compact and well separated are the clusters (maximum score = 1). The error bars indicate the CI estimated by repeatedly performing K-means clustering on randomly down-sampled versions of the original cohort (down-sampling to approximately 80% original cohort size), when keeping K fixed. The highest silhouette score is achieved for K = 3 clusters. Figure 5B depicts the data distribution along the first two principal components and corresponding cluster memberships.
We next examined the distribution of the estimated burden of each subtype in each cluster and the clinical outcomes. The clinical characteristics of the cases analyzed in this study are provided in Table 1. See Supplemental Table S5 for all the clinical metadata. Cluster 3 had significantly higher levels of the vessel-rich regions, whereas cluster 2 had significantly higher tumor necrosis relative to the rest of the cohort and slightly elevated levels of the chondroblastic subtype (Figure 5C). Overall, we observe that dogs belonging to cluster 3 had significantly worse clinical outcomes compared with the other two clusters. Figure 6A shows a Kaplan-Meier plot depicting differences in overall survival rates between dogs belonging to cluster 3 and rest of the cohort (log-rank test P = 0.038), whereas Figure 6B depicts the differences in disease-free interval rates between the dogs belonging to cluster 3 and rest of the cohort (log-rank test P = 0.0071). All dogs belonging to cluster 3 relapsed within 12 months after receiving adjuvant treatment. This negative association remained significant despite adjusting for relevant clinical parameters such as tumor location (proximal humerus versus non-proximal humerus), alkaline phosphatase levels (elevated versus normal), age, weight, sex, and adjuvant treatment type in a multivariable Cox proportional hazards regression model.
Table 1Clinical Characteristics of the Dog Osteosarcoma Cohort (N = 306)
Disease-free interval, time from surgery, days
Overall survival, time from surgery, days
Standard of care
Standard of care + sirolimus (rapamycin)
For continuous variables, values in parentheses represent the minimum and maximum range, and values outside the parentheses represent the median over the entire cohort. All other data are given as number (percentage).
Finally, we performed subgroup analysis to ensure prognostic signatures remain significant in unlabeled data not used in training. The first subgroup consists of 55 reviewed cases (n = 95 pathologist-annotated slides). The second subgroup consists of the remaining 251 unreviewed cases. In each subgroup, the survival association remains consistent, thus demonstrating the clinical utility of model predictions beyond cases previously annotated by the pathologist (Supplemental Figure S2).
Through the activities of the National Cancer Institute COTC, this study examines the largest data set of canine osteosarcomas to date for which complete clinical outcome data are available and standardized therapy was applied (n = 306). With the help of this large resource, we demonstrate how deep domain adversarial learning can be used to train a histologic subtype classifier that adapts from dog to human osteosarcoma despite utilizing a small fraction of human data for training. Although this is not the first application of deep learning in osteosarcomas,
it is the first attempting to identify histologic features of osteosarcoma that transfer from canine to human samples, to the best of our knowledge.
With the help of the trained species-agnostic histologic subtype classifier, we performed an unsupervised exploratory analysis of whole slide imaging data of 306 dogs and identified distinct clusters that respond differently to standardized chemotherapy based on the classifier-estimated burden of histologic subtypes. Our results are consistent with some prior reports indicating that the presence of specific histologic subtypes may have prognostic value
; however, a rigorous quantitative evaluation of OS histology that takes tumor heterogeneity into account has not been previously explored, likely because of the difficulty in accumulating a large enough data set and the immense manual labor by the pathologist in annotating each region. This is the first exploratory study using AI to define prognostic value of variant histologic features within a large population of dogs receiving standardized care in a prescriptive clinical trial. As with the diagnostic and therapeutic approach to any cancer, many separate factors should be considered when devising a treatment and prognosis. The predictive value of our approach should be considered alongside other patient factors and not considered the sole method by which prognosis can be assigned for canine patients with OS. Nevertheless, information gleaned from our approach is of substantial clinical value to clinicians treating dogs with OS.
In this study, we refrain from quantifying overlap between pathologist annotations and AI predictions using Dice or IoU metrics. These metrics are preferable in segmentation applications, where the ground truth segmentation boundaries are precisely defined.
Hence, it is not feasible for pathologists to precisely mark region boundaries of each histologic subtype at high resolution for each slide. Although the pathologist annotated most tumor tissue in all annotated sections, there are examples where unannotated tumor tissue was present. Interestingly, these cases offer another example demonstrating the ability of the model to identify tumor tissue that would not be captured by Dice or IoU metrics. For example, in Figure 2D, there are several regions that were predicted to contain osteoblastic tumor cells. On review, the pathologist was able to confirm the presence of osteoblastic tumor tissue in these locations (Supplemental Figure S3). This highlights a potential utility of AI in identifying foci of tumor distal to the main tumor mass. This may be particularly important in tumors that require complete excision and could help by re-orientating the pathologist toward specific regions to review.
In this study, tumors enriched for VR regions were associated with reduced disease-free interval and OS. These vascular structures define the rare telangiectatic subtype of osteosarcoma, which is characterized by blood-filled cystic spaces surrounded by thin septa lined by tumor cells.
suggested that telangiectatic OS carries a poor prognosis in human patients, others suggest that although there may be a correlation with clinical features, such as pathologic fracture, an association with prognosis is less clear.
(n = 45). In our case set, we defined VR regions as containing blood-filled spaces lined by tumor cells. On hematoxylin and eosin staining, these vascular spaces were multifocally lined by polygonal cells rather than flat, spindle-shaped cells, which were more likely to be interpreted as endothelium histologically. CD31 immunohistochemistry staining confirmed the presence of vessels lined by tumor cells in VR-annotated canine osteosarcomas (Supplemental Figure S2). Some VR regions also contained cellular debris, which has been described in human OS.
Although VR morphology was uncommon in our data set, the presence of tumor cell–lined vascular structures in largely solid tumors suggests that vascular differentiation can occur within a focal region of these histologically diverse tumors. Such tumors are less likely to be classified as telangiectatic OS, which may inhibit the prognostication of histologic subtype in OS. This is emphasized by a study of OS originating in the ulna (n = 30) that identified reduced survival in dogs with either pure or mixed telangiectatic morphology (ie, telangiectatic or osteoblastic-telangiectatic
This underlines the utility of AI, which allows pathologists to rapidly quantify the abundance of major and minor histologic patterns within heterogeneous tumors.
Despite the merits of this study, there are still a few notable limitations that should be considered. First, we did not have access to human clinical outcome data to assess the prognostic value added by our approach over what is currently clinically practiced for humans. A future direction will be to apply this method to a larger set of human OS images with matched clinical outcomes to determine algorithm performance in a translational setting. Second, our study is based on annotations from a single anatomic pathologist. Agreement between pathologists can vary based on the feature of interest. This may be greater in cases where pathologists must consider an aggregate of histologic features to assign a tumor grade. For example, in one veterinary study of osteosarcomas, agreement was considered moderate for necrosis (ICC = 0.626), whereas agreement on grade was fair using three different classification systems.
In the future, we aim to convene a comparative pathology board of M.D. and D.V.M. pathologists to review canine and human osteosarcoma histology with the goal of assessing the impact of our model on interobserver variability, identifying additional features, such as immune cell infiltration, that may be incorporated into our prognostic model alongside ongoing genomic work. Third, the data are severely imbalanced, with only a handful of canine and human tumor cases exhibiting uncommon histologic subtypes. To ensure that there exist enough training examples of each class for the patch-level classifier, pathologist-annotated whole slide images were broken into nonoverlapping patches scanned at high magnification and split at random into train validation and test sets (see Materials and Methods). Patch-based training of neural networks in digital pathology has enabled accurate detection and quantification of complex histologic features on few whole slide images because of thousands of image patches that can be extracted during training at high magnifications.
In this work, we reasoned that an adversarial learning approach could help neural networks overcome the bias that would be present in domain-specific training paradigms. Adversarial training can, however, be complex in practice compared with standard supervised learning approaches. This is especially relevant during initial phases of training, where noisy signals from the domain classifier can derail the learning algorithm.
This issue is mitigated by having a good initialization of model parameters and by gradually increasing the influence of domain classifier in the learning process, as defined in detail in Materials and Methods. Last, no additional manual quality control of surgical tumor specimens was completed before data collection from different sites. Instead, our model was adversarially trained to classify nontumor regions in addition to the six different histologic subtypes of osteosarcoma based on pathologist annotations. We expect the robustness and accuracy of the classification model to improve as additional data are collected.
In summary, deep domain adversarial learning could be a powerful addition to the modern pathologist's toolbox for identification of domain-agnostic histologic and molecular features of tumors and is likely to be useful for many other comparative oncology applications, especially where human data are scarce.
We thank the Comparative Oncology Clinical Trials Consortium (COTC) members for execution of the COTC-21/022 trials, which provided the clinical outcome data that were analyzed herein; and Dr. Markku Miettinen for granting access to 39 human osteosarcoma slides from his residency training materials.
Relationship between cluster membership and survival outcomes among previously annotated (A and B) versus unannotated cases (C and D). P values determining the significance of differences in survival rates were determined by the log-rank test. DFI, disease-free interval.