Abstract 2455: AI-based robust testicular cancer triaging, a promise for an efficient and cost-effective diagnostic/predictive modality

Testicular cancer is one of the most common malignancies in males aged 20-44, with its two major subtypes being Seminoma and Non-Seminomatous Germ Cell Tumors (NSGCT). Accurate classification is essential for patient management and treatment, after initial histologic diagnosis on radical orchiectomy specimens or biopsies of metastases. Histopathological diagnosis, though effective, is time-consuming, costly, and subject to inter-observer variability. Digital pathology combined with artificial intelligence (AI) promises to improve diagnostic consistency and efficiency while benefitting settings without expertise in these tumors. We seek rapid and robust triaging of these subtypes through a foundation model interrogating H&E-stained tissue sections and an attention-based Multiple Instance Learning (abMIL) approach providing interpretable predictions highlighting decision-driving regions. Retrospective multi-institutional data from 144 patients (254 slides) were used to develop our AI model. 118 patients (202 slides) were used for model training, 14 (30 slides) for validation, and 14 (22 slides) as an independent hold-out set. The UNI foundation model was pre-trained on diverse datasets and fine-tuned using abMIL, identifying discriminative features from the training data without requiring pixel-level annotations. Independent hold-out set evaluation revealed Area Under the Curve=96%, Accuracy=94.54%, Specificity=92%, and importantly, Sensitivity=100%. Specifically, all predicted NSGCT are correctly identified as NSGCT, and all Seminomas are always detected. In other words, no Seminoma case is overlooked, while 8% of the actual NSGCT may be misclassified as Seminoma. A sensitivity of 100% supports our model as a triaging tool, enabling faster decision-making for high-risk patients. Our proposed abMIL approach leveraging the foundation model UNI, provides a perfectly sensitive and scalable solution for clinical triaging between Seminoma and NSGCT testicular tumors in digitized tissue sections, ensuring subtype detection and supporting pathologists in improving diagnostic accuracy and efficiency. Furthermore, the model’s focus on highlighting AI decision-driving regions of the tissue section, renders it well-suited for large-scale clinical data, where it can make further data-driven contributions in extending our current disease knowledge. Citation Format: Siddhesh Thakur, Ujjwal Baid, Andres Martin Acosta, Muhammad T. Idrees, Michael Feldman, Spyridon Bakas. AI-based robust testicular cancer triaging, a promise for an efficient and cost-effective diagnostic/predictive modality [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 2455.