Comparative Analysis of Deep Learning Approaches for Histopathology-Based Survival Prediction in Hepatocellular Carcinoma

Background: Accurate prognostic stratification in hepatocellular carcinoma (HCC) remains challenging due to substantial tumor heterogeneity. Deep learning (DL)-based approaches, particularly convolutional neural networks (CNNs), have been widely applied for prognostic prediction using histopathological images. Recently, pathology foundation models combined with multiple-instance learning (MIL) have emerged as a promising alternative. In this study, we aimed to provide a comparative evaluation of conventional CNN-based and foundation model-based approaches for HCC prognosis prediction, with a focus on tissue selection strategies and cross-dataset generalizability. Methods: We compared a patch-level CNN approach with a foundation model-based MIL approach for overall survival (OS) prediction from hematoxylin and eosin-stained whole-slide images. A total of 256 patients from Seoul St. Mary’s Hospital (SSMH) and 334 patients from the TCGA dataset were included. Models were trained using either all tissue regions or tumor-only regions under five-fold cross-validation. Model performance was evaluated using the concordance index (C-index), Kaplan–Meier analysis, and time-dependent receiver operating characteristic analysis. Cross-dataset validation and combined-dataset training assessed generalizability. Results: In the SSMH dataset, the CNN model performed better with tumor-only regions (C-index 0.8308) than with all tissue regions (0.7498). In contrast, the foundation model-based MIL approach showed stable performance regardless of input regions (C-indices: 0.8701 and 0.8752). Similar stability was observed in the TCGA dataset (0.7744 and 0.7722). Cross-dataset validation showed reduced performance, indicating limited generalizability. Combining datasets did not lead to performance improvement. Subgroup analyses showed prognostic information beyond histologic grade, and feature visualization revealed relevant histopathologic patterns. Conclusions: A foundation model-based MIL approach provides robust and interpretable prognostic modeling in HCC. DL-based prediction offers complementary information beyond conventional clinicopathological variables. Future efforts integrating multimodal data and improving generalizability will be essential for clinical translation.