Abstract 2466: Pathomics prediction of thymic epithelial tumors histological subtypes: the AI-TET study

Abstract Background: Diagnosis of thymic epithelial tumors (TETs) is challenging due to multiple primary classes (A, AB, B1, B2, B3, C), based on tumoral and non-tumoral components, and their rarity. We aim to train and test a multiple-instance learning (MIL) model to classify the histological subtypes using hematoxylin-eosin+/-safran (HE/HES)-stained slides collected in the French RYTHMIC network. Patients and Methods: HE/HES slides from TET (A, AB, B1, B2, B3, C) cases addressed to RYTHMIC between 2012 and 2016 were retrieved and digitized. All pathological diagnoses were revised by central pathologists within the network. A MIL model was trained on digitized images, with no adjoint clinical features; internally validated using 3-repeated 2-fold cross-validation; further tested on a set of prospectively digitized cases addressed to RYTHMIC (and centrally revised) between 2022 and 2024, to classify the main TET types. Class predictions were evaluated by AUC scores and ROC curves. Interpretability was performed via Shapley value and heatmaps. Results: The training set comprised 456 slides from unique histological samples, with 243 (53%) obtained through thymectomy. Internal 3-repeated 2-fold cross-validation showed a mean AUC of 0.94 [std 0.005] for the overall prediction of subtypes. High-attention regions across slides contain both epithelial and lymphocytes, in different percentages according to subtype, showing the biological consistency of the prediction. In the test set (n=43), class prediction had a mean AUC of 0.96 [0.92-0.98]. Sensitivity/specificity for each class in the two sets are shown in the table. Conclusion: Our model holds promise as a diagnostic tool for TET, highlighting the potential of digital pathology in diagnosing rare and challenging entities. These results will be validated on a larger scale, and mixed components will be integrated. Internal cross-validation subtype(n) Sensitivity Specificity Test set subtype(n) Sensitivity Specificity A (33) 0.76 0.95 A (4) 0.25 1.00 AB (129) 0.74 0.95 AB (20) 0.85 0.87 B1 (26) 0.81 0.91 B1 (3) 0.33 1.00 B2 (110) 0.65 0.97 B2 (11) 1.00 0.81 B3 (60) 0.73 0.97 B3 (2) 1.00 1.00 C (98) 0.88 0.96 C (3) 0.67 1.00 Citation Format: Lodovica Zullo, Mathilde Bateson, Jose Carlos Benitez, Audrey Lupo, Damien Sizaret, Alvaro Lopez-Gutierrez, Daniela Miliziano, Kathryn Schutte, Juan David Florez-Arango, Pascale Missy, Vincent Thomas De Montpreville, Mercier Olaf, Jordi Remon, David Planchard, Nicolas Girard, Thierry Molina, Benjamin Besse. Pathomics prediction of thymic epithelial tumors histological subtypes: the AI-TET study [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2466.