Abstract

Computational pathology has seen significant advancements through the application of deep learning to Whole-Slide Images (WSIs). However, most existing models operate under a mono-tissue assumption, using only a single WSI to inform the prediction. This limitation is primarily driven by the structure of publicly available datasets, which rarely provide more than one slide per patient. Consequently, multi-tissue data integration remains underexplored in computational pathology, not due to a lack of clinical relevance but rather to a shortage of both public datasets and modeling strategies capable of handling multi-sample input. To address this gap, this study introduces MuGNet, a novel graph-based deep learning framework that enables patient-level prediction from a variable number of WSIs.

The MuGNet framework represents each patient as a network of tissue samples, with each WSI encoded as a node in a patient-specific graph. Edges are established to capture structural or biological relationships between tissues. A Graph Neural Network (GNN) with an attention-based message passing mechanism is then employed to process the patient graphs and generate patient-level predictions. MuGNet was evaluated on a private dataset of HGSC patients and investigated across three distinct prediction tasks to assess its robustness: one classification problem (Homologous Recombination Deficiency status prediction—HRD) and two regression problems (Overall Survival and Platinum Free Interval prediction—OS and PFI—).

MuGNet’s performance was compared against three state-of-the-art Multiple Instance Learning (MIL) models: ABMIL, DSMIL, and TransMIL. For the HRD status prediction task, MuGNet achieved an AUC of 0.7220±0.0739, which is approximately 6% higher than that of the most performant MIL model in this task. For the OS and PFI survival tasks, the model demonstrated an improvement in the C-index of approximately 10% over ABMIL and DSMIL, and up to 2% over TransMIL, highlighting its superior performance and robustness in handling multi-sample input. Beyond its predictive accuracy, MuGNet provides interpretability. The attention weights learned across the graph nodes allow for the identification of tissue sites that contribute most to the prediction, revealing that the model consistently prioritizes samples from metastasisrich regions, such as the omentum and lymph nodes.