Rhabdomyosarcoma (RMS) is the most common malignant soft tissue tumor in children, occurring at an incidence of approximately 4.5 cases per million children, or roughly 350 new cases per year in the United States. With a propensity for myogenic differentiation, RMS is classified into four histologic subtypes: embryonal (eRMS), alveolar (aRMS), spindle cell sclerosing (scRMS), and pleomorphic (adult type). These subtypes carry meaningfully different prognoses and require different treatment approaches, making accurate histologic classification a clinical priority rather than just an academic exercise.
Why subtyping is hard: The prognostic significance of RMS histology has been validated over decades, but the definitions have shifted over time. A long-standing debate in the field concerns whether aRMS diagnosis requires more than 50% alveolar features or merely any evidence of alveolar features at all. Adding further complexity, the presence of a FOXO1 gene fusion may be a stronger predictor of adverse outcome than histologic appearance alone in some cases. Other features, such as anaplasia, previously considered an adverse prognostic marker, may instead function as a surrogate marker for TP53 mutations rather than as an independent prognostic driver. These nuances underscore how much expert judgment is required, and how much interobserver variability can affect diagnosis.
The opportunity for deep learning: Whole slide image (WSI) scanning technologies now produce high-resolution digital tissue images that preserve cellular and subcellular detail. In parallel, advances in convolutional neural networks (CNNs) have enabled automated classification of complex visual patterns in pathology images, with demonstrated success in breast cancer grading, lung cancer mutation prediction, and lymph node metastasis detection. RMS had received almost no attention from the computational pathology community prior to this study, largely because of the rarity of the disease and the resulting scarcity of labeled training data.
This paper, published in the American Journal of Pathology in 2022, presents the first comprehensive computational study combining an RMS histology subtype classification model with a subsequent prognostic prediction model, both built on digitized hematoxylin and eosin (H&E)-stained slides from Children's Oncology Group (COG) clinical trial participants.
The dataset originated from a Children's Oncology Group (COG) approved biobanking study, ARST18B4-Q, titled "Biomarker Development through Digital Histology Evaluated by Automated Learners in Combination with Genomic Characterization of Rhabdomyosarcoma." Slides were prospectively collected from pediatric RMS patients enrolled in various COG clinical trials, with each patient contributing a single digitized whole slide image from a diagnostic biopsy taken at time of enrollment. In total, 340 WSIs were reviewed and classified by a board-certified pathologist (E.R. Rudzinski, Division of Anatomic Pathology, Seattle Children's Hospital), who provided the ground-truth subtype labels used for model training and evaluation.
Inclusion and exclusion criteria: Cases with anaplasia variants, sclerosing patterns, and rare mixed or epithelioid RMS patterns were excluded to maintain uniform within-class appearance during this exploratory study. This reduced the working cohort to 272 patients with clear, morphologically consistent subtype labels: 66 aRMS (24.3%), 196 eRMS (72.1%), and 10 scRMS (3.7%). No pleomorphic RMS cases were present, consistent with this being an exclusively pediatric dataset. The patient population was 68% male, with a median age at enrollment of approximately 6.4 years (2,321 days), and median follow-up of 1,830 days (approximately 5 years).
Clinical characteristics: The cohort included patients across all tumor stages, with 27.9% presenting with metastatic disease (T stage 4 and Surgical Group IV). Tumor sizes ranged from 0.6 to 27 cm (median 5.6 cm). FOXO1 fusion status was available for only 9.6% of patients (26 FOXO1-positive cases), reflecting the incomplete molecular annotation that is common in real-world pediatric oncology datasets. Clinical variables including sex, age, race, tumor stage and clinical group, tumor size, and FOXO1 fusion status were all available for each patient, enabling multivariate survival analysis.
The relatively small cohort size (272 patients across three subtypes, with only 10 scRMS cases) reflects the fundamental challenge of rare pediatric cancers: even large cooperative group studies accumulate cases slowly, limiting the volume of labeled training data available for machine learning development.
The RMS subtype classification model was developed using the InceptionV3 CNN architecture, selected for its favorable balance of parameter efficiency and classification performance relative to earlier architectures like AlexNet or VGG. InceptionV3 features parallel convolutional branches ("inception modules") that capture image features at multiple spatial scales simultaneously, which is well suited to histopathology where relevant features exist at both cellular and tissue-organization levels. The model was initialized with weights pretrained on ImageNet, a standard practice that provides a useful starting point for feature extraction even though natural images differ substantially from histopathology.
Patch extraction and annotation: From the 80 slides designated for model development (35 aRMS, 35 eRMS, 10 scRMS), each slide was manually annotated by a pathologist to identify five distinct tissue classes: aRMS tumor, eRMS tumor, scRMS tumor, other tissue (connective tissue, necrosis, glands), and white background. Image patches of 500 x 500 pixels at 20x magnification were extracted from each annotated region. The training set comprised 10,146 patches (3,107 aRMS, 3,793 eRMS, 1,246 scRMS, plus other tissue and background patches), the validation set 1,651 patches, and the patch-level test set 1,674 patches.
Data augmentation strategy: To counteract the small dataset size and the variability in tissue staining across slides from different institutions and time points, aggressive data augmentation was applied during training. Each image had a 70% probability of being augmented. Augmentation methods included Gaussian blur, random projective transformations (simulating slight changes in slide orientation), global color shifting (simulating staining intensity variation), per-channel color variation (simulating dye-specific inconsistencies), and random horizontal and vertical flipping. This augmentation pipeline substantially expanded the effective diversity of the training set without requiring additional patient samples.
Training used stochastic gradient descent (SGD) optimization in Keras with TensorFlow backend. The five-class model (aRMS, eRMS, scRMS, other tissue, background) was trained at the patch level and then applied to whole slides by sliding a 500 x 500 pixel window across the entire tissue area in steps of 500 pixels, generating a dense grid of patch-level predictions that could be assembled into a spatial heat map of tissue classification across the slide.
At the image patch level, the RMS classification model achieved an overall accuracy of 87.9% across all five tissue classes on the held-out patch-level test set of 1,674 patches. Broken down by class, patch-level accuracy was 84.0% for aRMS, 90.2% for eRMS, and 76.3% for scRMS. Accuracy for non-tumor tissues was 87.5%, and for the white background it reached 99.7%. The lower accuracy for scRMS (76.3%) reflects both the morphological similarity between scRMS and some eRMS patterns, and the very limited number of scRMS training cases (only 10 total slides, all used in training or validation).
Slide-level classification: The 192 slides in the slide-level test set (those not used in model development) were each processed by the trained classification model via the sliding window approach. Predictions from individual patches were aggregated to the slide level by identifying which RMS subtype class had the highest total patch count on the heat map. Of 31 aRMS slides, 29 were correctly classified (93.5%). Of 161 eRMS slides, 144 were correctly classified (89.4%). Receiver operating characteristic (ROC) analysis at slide level yielded an area under the curve (AUC) of 0.94 for aRMS classification and 0.92 for eRMS classification, both strong results for an automated pathology model applied to a rare disease with limited training data.
Heat map interpretation: The spatial heat maps generated for each slide showed that the model's predictions were anatomically coherent, correctly identifying RMS tumor regions as distinct from connective tissue and background. In representative examples shown in the paper, the model's predicted class regions aligned well with the pathologist's designations, with aRMS tumor regions highlighted in red and eRMS regions in orange. This spatial coherence is important because it suggests the model is learning genuine morphological features rather than slide-level artifacts like tissue processing or staining batch effects.
Because all 10 scRMS slides were used in model training and validation (given the extremely limited case count), scRMS slides were absent from the slide-level test set. This is a pragmatic decision given the rarity of scRMS, but it means that the slide-level classification performance metrics apply only to the aRMS versus eRMS distinction rather than the full three-class problem.
Building a prognostic model directly from scratch on a rare cancer dataset is particularly difficult, because survival-labeled cases are even scarcer than histologically labeled cases. Transfer learning addresses this challenge by reusing the image feature representations learned during the classification task and adapting them for a new predictive objective. In this study, the authors transferred the InceptionV3 network trained for RMS subtype classification to a new task: distinguishing eRMS patients by their survival outcome.
Defining high and low risk: The eRMS subtype was chosen for the prognostic model because it represented the largest available cohort (196 slides). From within this group, the authors defined low-risk patients as those with more than 8 years of progression-free survival without any event (relapse, second malignancy, or death), and high-risk patients as those who experienced disease progression during follow-up. Under these strict criteria, 30 low-risk and 30 high-risk eRMS slides were selected, with expert annotation identifying RMS tumor regions within each slide to ensure only tumor tissue contributed to prognostic predictions.
Network architecture for transfer: In the transfer learning model, all convolutional layers from the original InceptionV3 classification model were frozen, preserving the image feature extraction capacity learned during subtype classification. Only the final four layers (flattening, dense, dropout, and a second dense layer) were retrained on the eRMS outcome-labeled patches. This design forces the prognostic model to identify survival-relevant visual patterns within the feature space already organized by the classification task, rather than learning entirely new representations from the small prognostic training set. Training data comprised 2,483 image patches (1,337 from 24 low-risk slides, 1,146 from 24 high-risk slides), with 591 patches reserved for validation.
Critically, only image patches classified as RMS tumor by the subtype classification model were fed into the prognostic model. Connective tissue, necrosis, and background regions were excluded. This design prevents the prognostic model from learning features associated with tumor purity or stroma composition rather than intrinsic tumor biology, though it also means any stromal or microenvironmental signals are excluded from the analysis.
After training, the eRMS prognostic model was applied to the 136 eRMS slides not used in model development (the slide-level test set). Each slide was processed by the subtype classification model first to identify RMS tumor patches, then by the prognostic model to assign a patch-level probability of belonging to the high-risk category. A patch-level cutoff value was determined using the 60-slide model development set and applied without modification to the test set. Individual patients were classified as high risk if more than 50% of their tumor patches exceeded the high-risk probability cutoff, and as low risk otherwise.
Survival separation: Among the 136 test patients, the model classified 42 as low risk and 94 as high risk. Kaplan-Meier survival curves showed significantly different event-free survival between the two predicted groups (likelihood ratio test P = 0.02). The Cox proportional hazards model yielded a hazard ratio (HR) of 4.55 (95% CI: 1.04-19.96, P = 0.04) for the unadjusted comparison. This means predicted high-risk patients had approximately 4.5 times the event rate of predicted low-risk patients in the test cohort.
Multivariate analysis: The clinical relevance of the image-based risk prediction was further tested by adjusting for patient age and sex in a multivariate Cox model. After adjustment, the hazard ratio remained significant at 4.64 (95% CI: 1.05-20.57, P = 0.04). Neither age (HR 1.00, P = 0.78) nor sex (HR 1.04 for male, P = 0.94) contributed meaningfully to outcome prediction in this cohort, suggesting the image-based signal is the dominant predictor rather than a proxy for these clinical variables. This independence from standard clinical variables is an important finding, as it indicates the model may be capturing novel morphological information not encoded in routine clinical parameters.
It is worth noting that the confidence intervals are wide (1.05-20.57), reflecting the limited sample size of the test cohort (136 patients, divided into 42 low-risk and 94 high-risk). Wide CIs are expected in rare disease studies and do not invalidate the finding, but they do underscore the need for validation in larger, independent cohorts before clinical deployment.
Excluded RMS variants: Cases with anaplasia variants, sclerosing patterns, and rare epithelioid or mixed RMS subtypes were excluded from both the classification and prognostic models. These represent real patients encountered in clinical practice, and their exclusion means the published performance figures do not reflect the full diagnostic complexity a deployed tool would face. Anaplastic RMS variants in particular carry distinct prognostic implications, and any clinical-grade classification system would need to handle these patterns reliably. The authors acknowledge this as a direct limitation of the current exploratory study.
scRMS data scarcity: Only 10 scRMS slides were available, which is insufficient to develop or test a reliable classification model for this subtype. All 10 slides were consumed in model training and validation, leaving none for an independent test set. This means that the slide-level AUC results (0.94 for aRMS, 0.92 for eRMS) apply only to the binary aRMS-versus-eRMS distinction on the test set, not the full three-class problem. Real-world RMS pathology will encounter scRMS cases that the model has not been independently evaluated on at the slide level.
Single CNN architecture: Only InceptionV3 was evaluated, chosen based on prior experience rather than a systematic comparison. Alternative architectures (ResNet, EfficientNet, DenseNet, vision transformers) or ensemble approaches combining multiple models might yield better performance, particularly on the harder tasks of scRMS classification and prognostic prediction. The authors acknowledge this limitation and propose exploring additional architectures and ensemble methods as a future direction.
Prognostic model generalizability: The prognostic model's test set (136 patients) is small relative to the statistical precision required to draw robust conclusions, as evidenced by the wide confidence interval on the hazard ratio. The model was also developed and tested within a single dataset from COG, raising questions about whether the morphological features it learned will generalize to slides prepared at different institutions using different protocols and scanners. External validation on an independent institutional cohort has not yet been performed.
Improving classification coverage: The authors identify extending the classification model to include anaplastic variants and other rare RMS patterns as a primary near-term goal. This will require accumulating additional annotated cases, likely through continued expansion of the COG biobanking study and possible collaboration with international pediatric oncology groups such as the European Paediatric Soft Tissue Sarcoma Study Group (EpSSG) and the Cooperative Weichteilsarkom Studiengruppe (CWS). Larger multi-institutional datasets would also enable more rigorous external validation, which is the standard prerequisite for clinical translation of AI pathology tools.
Extending prognosis models to other subtypes: The current prognostic model is limited to eRMS because it is the most prevalent subtype and was the only one with sufficient survival-labeled cases for model development. As more aRMS and scRMS outcome data accumulate, extending transfer learning-based prognostic models to these subtypes becomes feasible, and may be particularly valuable for aRMS given its worse prognosis and the ongoing debate about optimal risk stratification criteria in FOXO1-positive versus fusion-negative cases.
Multi-modal integration: The authors highlight integration of genomic data as a key future direction. Combining histopathological image features with molecular profiles (FOXO1 fusion status, TP53 mutation, gene expression, copy number alterations) could produce prognostic models substantially more powerful than either data type alone. The growing availability of matched histology and genomic data from large cooperative group biobanks like the COG makes this an increasingly realistic goal. Multi-modal architectures such as cross-attention transformers that can fuse image features with tabular molecular data are an active research area in computational oncology.
Clinical utility and workflow integration: Beyond prediction accuracy, the path to clinical implementation requires prospective validation in real diagnostic workflows, clinician acceptance and interpretability, regulatory approval, and integration with digital pathology platforms. The authors frame the current models as potential aids to pathology evaluation and additional parameters for risk stratification, rather than replacement tools, which is an appropriate framing for the current stage of evidence. Demonstrating that AI-assisted reading reduces subtype disagreement rates between community and expert pathologists would constitute compelling clinical evidence for adoption.