Abstract 2423: Automated classification of pediatric sarcoma using digital histopathology

Introduction: Pediatric sarcomas are challenging to accurately classify due to their rarity and the wide diversity of subtypes. The process requires highly specialized pathologists as well as molecular and genetic testing that is expensive, takes time, and is not universally available. Deep neural network models (DNNs) trained on histopathology slides can reduce the time and cost to diagnosis and attenuate disparities in care based on geographical location and socioeconomic status. Here, we demonstrate the efficacy of automated image analysis for assigning sarcoma diagnoses across centers by identifying Ewing Sarcoma (ES), distinguishing rhabdomyosarcoma (RMS) vs non-rhabdomyosarcoma soft tissue sarcomas (NRSTS), as well as classifying alveolar, embryonal, and spindle cell RMS subtypes. Methods: Images were collected from Massachusetts General Hospital, Yale School of Medicine, St Jude Children’s, and the Children’s Oncology Group. A total of 691 images were collected across all centers, including 9 different sarcoma subtypes. Images were harmonized using our published STQ pipeline including focus checking, resolution standardization, and stain normalization. Image tiling and feature extraction was performed comparing multiple deep learning backbones (CTransPath, UNI, CONCH). Tile-level features were transposed to whole slide-level representations using our published SAMPLER method in which each feature is represented as the vector of decile values of its distribution across all tiles. Resulting feature sets were fed into logistic regression models for sarcoma classification tasks. We benchmark this approach against transformer based multi-head self-attention models trained on a V100 GPU. Results: We are able to distinguish ES from all other sarcoma types with an AUROC of 0.966. In the task of NRSTS v. alveolar RMS v. embryonal RMS we achieve an AUROC of 0.939. Restricting to RMS subtypes, we distinguish alveolar from embryonal with an AUROC of 0.95. Also, despite uneven sample representation, we obtain an AUROC of 0.88 for alveolar v. embryonal v. spindle type RMS. Finally, we have developed spatially-resolved attention maps, which provide interpretability for the regions of a slide that contain malignant cells. Conclusion: To our knowledge, we have amassed the largest multicenter pediatric sarcoma imaging dataset with broad representation across subtypes, anatomical locations, race, and sex. Our pipeline allows for further integration of images from any center, which may be co-analyzed to attenuate center-specific batch effects. Our classification accuracies are state of the art for multiple tasks fundamental to clinical sarcoma pathology, including novel multiclass distinction among three different RMS subtypes. Importantly, our pipeline and SAMPLER model can be run with minimal computational requirements, allowing for broad accessibility. Citation Format: Adam Thiesen, Sergii Domanskyi, Ali Foroughi pour, Jingyan Zhang, Todd B. Sheridan, Steven B. Neuhauser, Alyssa Stetson, Katelyn Dannheim, Danielle B. Cameron, Shawn Ahn, Hao Wu, Emily R. Christison-Lagay, Carol J. Bult, Jeffrey H. Chuang, Jill C. Rubinstein. Automated classification of pediatric sarcoma using digital histopathology [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular s); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1): nr 2423.