Abstract 2450: Slide-level foundation models exhibit effective diagnostic and prognostic performance in H&E-stained histopathology

Abstract Introduction: Computational pathology has progressed from tile-level to slide-level foundation models, eliminating the need to train separate tile aggregation networks. Slide-level foundation models are trained on extensive datasets of whole-slide images. For instance, Prism is trained on 587, 196 slides from 17 tissue origins, while GigaPath is trained on 171, 189 slides covering 31 tissue types, both sourced from private data cohorts. By leveraging vast amounts of data to capture intricate tissue patterns, slide-level models promise to extract clinically relevant features. However, currently there is limited independent research evaluating the performance of these models in more complex tasks, such as multiclass sub-typing diagnosis and survival predictions. Methods: We curated FFPE H&E slides from both the public TCGA and our in-house cohort. From TCGA, we collected slides from 5 cancer types: 344 BLCA, 505 KIRC, 348 LIHC, 350 STAD, and 500 THCA cases. The in-house cohort consisted of rare cancer cases, categorized into 54carcinoma, 67 sarcoma, and 18 melanoma cases. We applied slide-level foundation models, Prism and GigaPath, to embed the slides into feature representations. For the diagnosis task, we used a support vector machine (SVM) model to classify TCGA slides into 5 categories and the in-house slides into 3 categories. For the prognosis task, we trained a multi-layer perceptron (MLP) model to predict the hazard risk. We conducted 5-fold cross-validation for both diagnosis and prognosis tasks. Results: On TCGA slides, we achieved a classification accuracy of 0.973±0.008 using Prism and 0.944±0.022 using GigaPath. For the in-house slides, the classification accuracy was 0.885±0.042 with Prism and 0.848±0.056 with GigaPath. These results demonstrate the strong discriminative capabilities of the feature representations extracted by both slide-level foundation models. Notably, Prism outperformed GigaPath on both the TCGA and in-house datasets. For survival prediction, we applied the Prism model to five TCGA cancer types and the in-house cases. The C-index values were 0.595±0.029 for BLCA, 0.719±0.039 for KIRC, 0.655±0.058 for LIHC, 0.591±0.046 for STAD, 0.720±0.114 for THCA, and 0.592±0.025 for all in-house cases combined. These varying levels of survival prediction performance reflect the challenges posed by different tasks, while the lowest value being above 0.59 suggests prognostic information was adequately extracted. Conclusion: By evaluating the diagnostic and prognostic capabilities of slide-level foundational models, we demonstrated their effectiveness on both public and in-house datasets. Our findings corroborate that these models can distill clinically relevant feature representations even without fine-tuning, underscoring their potential for broad computational pathology applications. Citation Format: Pingjun Chen, Alejandra G. Serrano, Tanishq Gautam, Sara Ranjbar, B. Leticia Rodriguez, Patient Mosaic Team, Luisa M. Solis Soto, Yinyin Yuan. Slide-level foundation models exhibit effective diagnostic and prognostic performance in H&E-stained histopathology [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 2450.