Liver Segmentation using Deep learning with Multi-Core Pooling Modules

Abstract

Accurate liver segmentation in laparoscopic procedures is an important challenge owing to low contrast, non-uniform illumination, occlusion from instruments, and non-uniform organ boundaries. Inaccurate or delayed segmentation may enhance surgical risk, operating time, and recovery time for the patient. Although convolutional neural networks (CNNs) have shown good performance in medical image segmentation, previous approaches either obtain high accuracy at the expense of computational efficiency or obtain lightweight operation with low precision but usually lose fine-grained boundary details. To overcome the above shortcomings, we introduce a new deep learning network for laparoscopic liver segmentation that combines an InceptionV3 backbone with Multi-Core Pooling (MCP) and an enhanced Atrous Spatial Pyramid Pooling (ASPP) module. The proposed hybrid architecture extracts multi-scale contextual information, preserves boundary precision, and is computationally light enough for future real-time surgery use. The model is tested on the publicly available M2CAI Segmentation dataset and exhibits better performance than state-of-the-art algorithms like U-Net++, nnU-Net, and SwinD-Net. Our findings reveal that the proposed model provides robust, accurate, and efficient liver segmentation, presenting a promising solution for real-time intraoperative guidance.