MRNet: A Multi-Resolution Dual-Task Framework for Micrometastases Detection in Breast Cancer Sentinel Lymph Nodes

Deep learning algorithms for detecting micrometastasis in breast cancer lymph nodes show promising results when used complementarily to improve the efficiency of pathologists' routines. However, in the current literature, there are critical limitations, poor performance on isolated tumor cells (detection rates<40%), and high false-positive rates due to confusion with nerves or contamination. In this article, we introduce MRNet, a novel multi-resolution dual-task framework that addresses these challenges through task-optimized resolution processing and annotation imprecision handling. Unlike existing approaches that use uniform resolution, our method employs high-resolution patches (level 0) for classification to capture subtle micrometastatic features, while using moderate resolution (level 3) for segmentation to mitigate annotation imprecision inherent in histopathological datasets. Our preprocessing pipeline enables efficient processing of gigapixel whole-slide images without resolution loss, while post-processing reconstruction maintains spatial coherence. We achieved state-of-the-art classification performance with area under the curve $=0.998$ , while reaching free response operating characteristic $=0.6124$ in localization tasks. Most significantly, our multi-resolution strategy demonstrates that the disconnect between patch-level and slide-level performance in existing methods can be systematically addressed through resolution-aware design.