Towards scalable cancer prognosis by aligning semi-supervised learning with survival modeling

Prognostic assessment via whole-slide images (WSIs) is a key task in computational pathology. However, the development of robust predictive models is often bottlenecked by the scarcity of survival-annotated data, given the longitudinal nature of clinical follow-up. Despite its potential, semi-supervised learning (SSL) remains underexplored in this field. Existing SSL methods, primarily designed for natural image classification, rely on the assumption of marginally independent categories, which makes them difficult to adapt to the high-dimensional complexity of WSIs and the temporal continuity inherent in survival outcomes. To bridge this gap, we propose a novel semi-supervised multiple instance learning (MIL) framework tailored for WSI-based prognosis. Specifically, we first introduce a morphology-perceived MIL network (Morph-MIL) that jointly models tissue composition and global semantic representations. To maintain a stable reference, a dynamic memory bank is employed to encapsulate the multi-level representations of labeled samples (censored & uncensored). Furthermore, we develop a pseudo-labeling strategy that assigns discrete labels containing temporal relationships to annotation-free data by aligning their features with the memory bank through bin-wise aggregation. Comprehensive evaluation establishes the superiority of the proposed framework, while extensive ablation analyses and visualizations confirm its efficacy in prognostic prediction and biological interpretation. Consequently, this study offers a scalable solution for WSI survival modeling in scenarios with limited annotations. The code is available at: https://github.com/yihuangswust-alt/SSL-Morph_MIL.