Pathomics-based prediction of thymic epithelial tumor subtypes within the French RYTHMIC network.

Abstract

8112 Background: Thymic epithelial tumors (TETs) are classified into primary subtypes (A, AB, B1, B2, B3, C) and mixed classes, determined by varying proportions of tumoral and non-tumoral components, and different subtypes have different prognosis. This heterogeneity, combined with their rarity, poses significant diagnostic challenges that impact tumor treatment. We aim to develop and test a multiple-instance learning (MIL) model capable of classifying TETs major histological subtypes from hematoxylin-eosin/hematoxylin-eosin-saffron (HE/HES)-stained slides. Methods: Cases who underwent central revision by a panel of expert pathologist between 2012 and 2016 in the context of the French national RYTHMIC network were retrospectively collected, and their HE/HES slides were digitized in whole slide images (WSIs), forming the training cohort. A MIL model was trained exclusively on digitized WSIs, without clinical features, and internally validated using 3-repeated 2-fold cross-validation for the classification of major TET subtypes: A, AB, B1, B2, B3, C. Prospectively digitized WSIs from the RYTHMIC network (2022–2024) served as the testing cohort. Class predictions were assessed using AUC scores and ROC curves. Interpretability was explored through Shapley values and heatmaps. Results: A total of 456 WSIs from unique histological samples formed the training cohort, with 243 (53%) samples obtained via thymectomy. The most represented subtype was AB (n=129, 28%), followed by B2 (n=110, 24%). Internal validation achieved a mean AUC of 0.94 [sd 0.005] for histological subtypes classification. High-attention regions identified on the slides featured varying proportions of epithelial cells and lymphocytes, consistent with the biological characteristics of each subtype. The test set comprised 75 WSIs from unique histological samples, with 63 (84%) obtained via thymectomy. The most represented subtype was AB (n=35, 47%), followed by B2 (n=19, 25%). In the test set, the model achieved a mean AUC of 0.89 [95%CI 0.83–0.93] for histological subtypes classification. Conclusions: Our model shows promise for diagnosing TET major subtypes, emphasizing the value of digital pathology in identifying and classifying rare entities. We are currently reviewing discrepancies between MIL and pathologists' diagnoses in subtype classification within the test set to evaluate the potential of artificial intelligence in aiding the diagnosis of complex cases. The final results will be presented at the congress.