IANet: Important-Aware Network for Microscopic Hyperspectral Pathology Image Segmentation

Microscopic hyperspectral pathology image (MHPI) provides a wealth of reference information for medical diagnosis. However, the accompanying high-dimensional complex features bring great challenges to the task of pathology image segmentation. In this paper, a novel important-aware network (IANet) for MHPI segmentation is proposed. IANet builds an encoder with pre-trained ResNet34 and hierarchical fusion pyramid (HFP) modules to extract multiscale high-level features in MHPIs. Furthermore, an important-aware fusion (IAF) module is developed and embedded in the skip connection to simultaneously highlight task-relevant salient spatial and semantic features. In particular, a target-aware edge enhancement (TAEE) module is designed to improve the edge segmentation effect of the target regions. The proposed IANet realizes the full mining of the intrinsic information in MHPIs, and has efficient feature fusion and fine edge segmentation capabilities. The experimental results show that the proposed method outperforms other state-of-the-art methods on the MHPI segmentation task, providing an effective way for auxiliary medical diagnosis.