Abstract
This article introduces a novel liver and gallbladder segmentation framework, named Adaptive Multi-Scale Attention YOLO Network (AMAY-Net), designed for semantic segmentation of laparoscopic cholecystectomy images. Building upon the powerful feature extraction capabilities of You Only Look Once (YOLO), AMAY-Net incorporates several advanced modules to enhance performance in medical image segmentation tasks. First, a multi-scale feature extraction module is employed to capture anatomical structures of various sizes, ensuring effective detection of large organs like the liver and smaller structures such as the gallbladder and surgical instruments. Second, an adaptive class-balancing loss function is implemented to dynamically adjust the weights of underrepresented classes, improving the segmentation accuracy of small structures. Additionally, the network integrates a spatial and channel attention mechanism, enhancing the focus on critical regions in the image. Finally, residual connections are introduced in the YOLO backbone to improve feature propagation and gradient flow efficiency. Experimental results demonstrate that AMAY-Net achieves superior performance on the CholecSeg8k dataset, with significant improvements in the segmentation accuracy of key anatomical structures such as the liver and gallbladder.