Multicenter Histology Image Integration and Multiscale Deep Learning for Machine Learning-Enabled Pediatric Sarcoma Classification

Pediatric sarcomas are rare and diverse, often leading to misclassification that hampers prognosis and treatment planning. We collected and harmonized histology images from multiple medical centers and developed accurate, generalizable classifiers for sarcoma subtype classification using deep learning techniques on digitized histology slides. A pediatric sarcoma histology dataset was amassed from three medical centers as well as the Children’s Oncology Group (COG), and a computational pipeline was implemented to address batch effects and standardize processing. Input parameter combinations (e.g. tile size, multi-scale feature extraction) were tested and optimized, and multiple convolutional neural network (CNN) and vision transformer-based (ViT) deep learning backbones were used as feature extractors for our SAMPLER whole slide image (WSI) representations, which we benchmarked to create an optimized, sarcoma-specific workflow. Models were trained for automated classification of rhabdomyosarcoma (RMS) versus non-rhabdomyosarcoma (NRSTS) and to further distinguish subtypes including alveolar, embryonal and spindle cell RMS as well as Ewing sarcoma. A total of 867 whole slide images were harmonized to generate a digital histology library with representation from 10 pediatric sarcoma subtypes. Across multiple classification tasks, ViT are superior to CNN models and multiscale feature sets consistently outperform single-scale models. Optimized classifiers accurately predicted RMS versus NRSTS (AUC 0.969±0.026) as well as alveolar versus embryonal RMS (AUC 0.961±0.021). Our two-stage classifier distinguished Ewing sarcoma from other NRSTS (AUC 0.929). Additionally, our models are computationally more lightweight than standard transformer implementations (model size 0.111 v. 1.9MB, training time > 3 orders of magnitude faster). Digital histopathology can successfully classify pediatric sarcomas, providing results that are reproducible, mitigate inter-observer bias, and can be implemented remotely. We demonstrate that image harmonization overcomes the pitfall of overfitting to single institutional data while maintaining state-of-the-art performance across multiple classification tasks. The developed classifiers provide a basis for more precise histology-based sarcoma diagnostics, enabling global access to improved prognostication and treatment planning. Pediatric sarcomas present diagnostic challenges due to their rarity and diverse subtypes, often requiring specialized pathology expertise and costly genetic tests. To overcome these barriers, we developed a computational pipeline leveraging deep learning methods to accurately classify pediatric sarcoma subtypes from digitized histology slides. To ensure classifier generalizability and minimize center-specific artifacts, we collected and harmonized a dataset comprising 867 whole slide images (WSIs) from three medical centers and the Children’s Oncology Group (COG). Multiple convolutional neural network (CNN) and vision transformer (ViT) architectures were systematically evaluated as feature extractors for SAMPLER-based WSI representations, and input parameters such as tile size combinations and resolutions were tested and optimized. Our analysis showed that advanced ViT foundation models (UNI, CONCH) significantly outperformed earlier approaches, and incorporating multiscale features can enhance classification accuracy. Our optimized models achieved high performance, distinguishing rhabdomyosarcoma (RMS) from non-rhabdomyosarcoma (NRSTS) with an AUC of 0.969±0.026 and differentiating RMS subtypes (alveolar vs. embryonal) with an AUC of 0.961±0.021. Additionally, a two-stage pipeline effectively identified scarce Ewing sarcoma images from other NRSTS (AUC 0.929). Compared to conventional transformer encoder architectures used for WSI representations, our SAMPLER based classifiers were more lightweight (0.111 MB vs. 1.9 MB) and three orders of magnitude faster to train. This study highlights that digital histopathology paired with rigorous image harmonization provides a powerful solution for pediatric sarcoma classification. Our models reduce inter-observer variability, augment diagnostic precision, and have the potential to increase global accessibility to robust diagnostics, improving time to diagnosis and subsequent treatment planning. Digitized H&E-stained histopathology slides offer a transformative opportunity to develop artificial intelligence-based models for pediatric sarcoma histological classification. We leveraged advanced imaging analysis and deep learning techniques to develop classifiers capable of accurately identifying and differentiating subtypes of sarcoma. This innovation paves the way for imaging-based diagnostic methods that can be universally applied, ensuring that patients receive precise and timely diagnoses regardless of their geographic location or the resources available at their treatment centers. Consequently, this approach has the potential to significantly enhance the equity of access to precision diagnostics, ultimately improving prognostics and treatment planning for pediatric sarcoma patients worldwide.