Artificial Intelligence-Based Histopathological Analysis to Assist Pathologists in Diagnosing Ewing Sarcoma and...

Ewing sarcoma is a highly malignant tumor whose histological appearance often overlaps with that of other undifferentiated round-cell and ovoid-cell tumors, making accurate diagnosis challenging. Selecting the most appropriate immunohistochemical and molecular tests is critical, particularly when only limited core biopsy material is available and tissue preservation for additional molecular or biomarker testing is required. Artificial intelligence (AI)-based histopathological image analysis is emerging as a promising diagnostic tool in oncology. This study evaluated the potential of AI-based histopathological analysis to assist pathologists in the morphologic classification of Ewing sarcoma and selected histological mimics using tissue microarrays (TMAs), rather than to identify or predict molecular alterations. We analyzed 1926 digitized histological cores, from 729 patients, assembled into 45 tissue microarrays. The dataset comprised 517 Ewing sarcomas (ESs), 367 rhabdomyosarcomas (RMSs), 187 chondrosarcomas (CHSs), 138 gastrointestinal stromal tumors (GISTs), and 124 synovial sarcomas (SSs). A weakly supervised multiple-instance learning (MIL) framework with transformer-based aggregation was developed using only core-level diagnostic labels. The model achieved classification accuracies of 97.1% for Ewing sarcoma, 80.0% for rhabdomyosarcoma, 85.7% for gastrointestinal stromal tumors, 80.0% for chondrosarcoma, and 76.0% for synovial sarcoma. The overall classification accuracy was 91.6%, with no misclassifications between Ewing sarcoma and rhabdomyosarcoma. These findings highlight the model's robustness in distinguishing tumor entities that present a well-recognized diagnostic challenge in routine pathology. The AI algorithm demonstrated strong potential as a diagnostic adjunct for assisting pathologists in the morphologic classification of Ewing sarcoma and selected tumor entities based on histopathological features. Although these tumor entities are characterized by specific molecular alterations, the model was not designed to identify or predict molecular alterations directly. Instead, it supports clinical decision-making by recognizing morphologic patterns associated with diagnostically defined tumor entities. The integration of AI-based histopathological analysis with immunohistochemistry and contemporary molecular diagnostic techniques has the potential to improve diagnostic accuracy, optimize the use of ancillary testing, enhance our understanding of genotype-phenotype relationships, and ultimately support more effective patient management.