Transem-Net: Transformer Based Efficient Multi-Magnification Network for Histopathology

Analyzing digital Whole Slide Images at multiple magnifications simulates the workflow of a pathologist and gives much-needed context from global tissue structures. Transformer models are especially adept at capturing long-range dependencies and naturally lend themselves well to multi-magnification tissue analysis. In this paper, we present TransEM-Net, an efficient transformer-based architecture that collates information from two encoders operating on concentric images with varying fields of view. We evaluate TransEM-Net on lung and prostate cancer tissue datasets to demonstrate that the model outperforms the baseline as well as contemporary architectures. At the same time, TransEM-Net also provides a better balance between performance and model efficiency.