Abstract B020: Automated classification of thymic epithelial tumors. A novel deep learning approach

Abstract Introduction: Accurate histological classification of thymic epithelial tumours (TETs), including subtypes A, AB, B1, B2, B3, and thymic carcinomas (TC), is essential for prognosis and treatment planning. However, expert-level classification remains challenging due to significant inter-observer variability and the rarity of these tumors. This study presents a novel deep learning approach for subtype classification using digital histopathology. Methods: Hematoxylin and eosin (H&E)-stained whole slide images (WSIs) from the Cancer Genome Atlas Program (TCGA) were used. The dataset included TETs resections from 119 patients. WSIs were divided into 224×224 pixel patches. We used a foundational model called UNI to extract high-dimensional features. UNI was pretrained on over 100 million histopathology images across 20 major tissue types. We then employed an attention-based Multiple Instance Learning (MIL) model to aggregate patch-level information for slide-level classification. The key innovation in our approach is the implementation of a biologically-informed hierarchical loss function with three components: a multiclass classifier to distinguish A/AB, B1–B3, and TC categories, a binary classifier to differentiate A and AB subtype, and an ordinal classifier to model the biological continuum among B1–B3 subtypes, implemented using a novel binary encoding scheme where each class receives one more "1" bit than the previous class (e.g., B1: [0,0], B2: [1,0], B3: [1,1]). This approach represents the biological continuum of B-subtypes based on increasing epithelial-to-lymphocyte ratios. Results: The model was evaluated using 3-fold cross-validation. Compared to random classification accuracy of only 17%, our model achieved an overall six-class classification accuracy of 59.4% (95% CI: 55.3–63.5) with a Cohen’s kappa of 0.485 (0.437–0.534). When collapsed into three high-level classes (A/AB vs. B1–B3 vs. TC), accuracy improved to 81.0% (76.5–85.5) with a Cohen’s kappa of 0.678 (0.598–0.757). Performance was exceptionally high for TC classification: Accuracy: 95.8% (92.5–99.0), Sensitivity: 94.4% (86.0–100), and Specificity: 96.0% (93.1–98.9). Conclusions: This deep learning approach demonstrates strong performance in classifying TETs subtypes, with especially high accuracy for thymic carcinomas. Prior studies have shown substantial inter-observer variability, with expert reviews at specialized centers resulting in a different histological classification for up to 56% of referred thymic tumor cases—potentially altering treatment decisions in more than 40% of these instances. Our model offers a promising tool to augment diagnostic accuracy and reduce variability, particularly in settings lacking expert pathology review. Citation Format: Matteo Sacco, Erica Pietroluongo, James M. Dolezal, Anna Di Lello, Mirella Marino, Alessandra Esposito, Maha AT. Elsebaie, Marina C. Garassino. Automated classification of thymic epithelial tumors. A novel deep learning approach [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B020.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.