MambaBack Framework Addresses Challenges in Whole Slide Image Analysis for Computational Pathology

A new research paper introduces MambaBack, a framework designed to overcome limitations in Whole Slide Image analysis for computational pathology. Whole Slide Image analysis is crucial for cancer diagnosis by examining morphological and architectural details across various magnifications. Multiple Instance Learning serves as the standard framework for this type of analysis. While Mamba has emerged as a promising backbone for Multiple Instance Learning, surpassing Transformers due to efficiency and global context modeling from Natural Language Processing, existing Mamba-based approaches face significant issues. These challenges include the disruption of 2D spatial locality when flattening to 1D sequences, inadequate modeling of fine-grained local cellular structures, and high memory consumption during inference on edge devices. Research such as MambaOut indicates that Mamba's State Space Model component is unnecessary for local feature extraction, where Gated Convolutional Neural Networks are sufficient. The MambaBack framework specifically addresses the need for both fine-grained local feature extraction and global context modeling in Whole Slide Image analysis. This work is documented in the preprint arXiv:2604.15729v1, which was announced as a cross-disciplinary study.