(P 177) DEEP LEARNING AND DIMENSIONALITY REDUCTION FOR SARCOMA SUBTYPING

Objective: Synovial sarcoma (SS) and leiomyosarcoma (LMS) are distinct soft tissue sarcomas with overlapping histologic characteristics, complicating accurate diagnosis. Traditional diagnostic methods involving histopathologic evaluation, immunohistochemistry, and molecular testing are often time-consuming, subjective, and resource intensive. The objective of this study was to leverage convolutional neural networks (CNNs) and Uniform Manifold Approximation and Projection (UMAP) to improve diagnostic accuracy, reduce reliance on ancillary testing, and provide interpretable deep learning visualizations.

Methods: This computational study used six SS and seven LMS whole-slide images from The Cancer Genome Atlas. After background removal, slides were segmented into 225×225-pixel grayscale images, and a random sample of 5000 of these images (2500 SS, 2500 LMS) was drawn due to limitations in computing power. Data were partitioned into training (80%), validation (10%), and testing (10%) sets. We employed a pre-trained ResNet50 model and used transfer learning with data augmentation, dropout layers, Adam optimization, and early stopping to mitigate overfitting. UMAP was applied to visualize high-dimensional CNN embeddings. Generalized model performance will be validated on independent samples in the future.

Results: Preliminary results demonstrated that CNN-based classification effectively distinguished SS from LMS (accuracy: 97.4%) capturing morphological subtleties not readily discernible through conventional methods. UMAP visualization facilitated interpretability by illustrating distinct clustering of CNN-derived features for SS and LMS.

Conclusion: Integration of CNNs and UMAP offers a promising AI-driven approach to improve sarcoma classification accuracy, potentially reducing reliance on ancillary diagnostic testing. Future work will address validation with external datasets to further establish clinical utility and generalizability.

Table 1: Model Performance Metrics