BPP-net: bio-inspired parallel pathway network for liver and tumor segmentation

Liver tumor segmentation, the intricate process of identifying and delineating tumor anatomical structures or pathological regions from medical liver images, plays a crucial role in assisting physicians with diagnosis and treatment planning. Automatic liver tumor segmentation is a highly complex task due to the low contrast of the target, its wide range of occurrences, and significant scale variations. Inspired by the intricate serial-parallel structure and information processing mechanism of the biological visual system, we propose the bio-inspired parallel pathway network (BPP-Net). In analogy to the parallel pathways of biological vision, BPP-Net consists of three parallel streams composed of antagonistic-Mamba modules at different scales, designed to extract target features across multiple scales. The antagonistic-Mamba module mimics the visual antagonistic properties of cells in the visual pathway, enhancing the network’s ability to extract features from low-contrast regions. Furthermore, inspired by the multi-channel interaction characteristics of the bio-visual IT layer, we propose a dual-channel feature fusion decoder to progressively refine segmentation results. We evaluate our method on the LiTS2017 and BEIDC2024 datasets, and the proposed BPP-Net achieves state-of-the-art (SOTA) performance with only 0.14 M parameters, demonstrating promising potential in the field of bio-computer vision.