Abstract

Early and accurate cancer detection through pathology imaging plays a critical role in improving patient outcomes. Despite promising advances from deep learning (DL) models, their real-world deployment is hindered by significant challenges.

To address the performance degradation caused by domain shifts—variations from different imaging devices, staining protocols, and patient demographics—we introduce PathTTT. This novel framework enhances model robustness at the patch level by combining Test-Time Training (TTT) with Model-Agnostic Meta-Learning (MAML). This bi-level optimization strategy leverages MAML to train a model for rapid adaptation, while TTT dynamically fine-tunes its parameters during inference, allowing for effective generalization to unseen distributions.

A separate, key challenge in computational pathology is the effective analysis of whole slide images (WSIs). While multi-instance learning (MIL) has become a widely adopted paradigm for WSI classification, it often suffers from overfitting due to the extremely high dimensionality and heterogeneity of WSIs, combined with the limited availability of annotated data. To address these challenges, we propose Masked Feature Embedding for Multi-Instance Learning (MFE-MIL), an innovative method designed to enhance existing end-to-end MIL pipelines. By introducing a masked feature embedding prediction task, our method provides a strong self-supervised signal that forces the model to learn highly discriminative and context-aware representations from instance features.


Extensive experiments on multiple benchmark pathology imaging datasets demonstrate that both PathTTT and MFE-MIL consistently outperform state-of-the-art methods in their respective domains. These results collectively underscore the potential of this work to facilitate more reliable and generalizable cancer detection systems in real-world clinical applications.