Pathology-guided contrastive pretraining enriches preoperative CT representations for prognosis

Abstract

PURPOSE: Prognosis prediction is important for personalized cancer treatment. Histopathology imaging can provide key prognostic information but it relies on tissue obtained during surgery, which restricts its use for preoperative decision-making. Preoperative imaging modalities such as computed tomography (CT) are more accessible but often lack the pathological context needed for accurate prognosis. Multi-modal approaches that combine imaging and pathology aim to bridge this gap, yet they typically require both modalities at inference time, making them impractical for preoperative decision-making. METHODS: To address this, we adopt a multi-modal contrastive learning (MMCL) framework that uses histopathology embeddings to guide CT feature learning. This approach enables the CT encoder to capture pathology-informed patterns during pretraining. At inference, MMCL uses only CT and routine clinical factors, allowing practical preoperative application. We evaluate MMCL on the MMIST-ccRCC dataset using cross-validation for predicting 12-month survival status. RESULTS: MMCL achieves an AUROC of 76% using only preoperative CT and clinical factors during inference, surpassing state-of-the-art multi-modal models that leverage both preoperative data and postoperative pathology. Even with CT alone, MMCL matches the performance of state-of-the-art multi-modal baselines, underscoring its robustness. CONCLUSION: Our findings demonstrate that contrastive learning allows CT encoders to extract richer and prognostically relevant features. Importantly, at the testing stage, more accurate and non-invasive survival prediction is achieved using only preoperative CT and general clinical features. This approach provides a promising strategy to improve preoperative prognostic assessment, supporting early treatment planning and guiding personalized cancer care.

Ähnliche Arbeiten