Path-enhanced chunking approach with residual attention for medical image segmentation

Medical image segmentation is an essential component of computer-aided diagnosis. While U-Net has been widely used in this field, its performance can be limited by incomplete feature information transfer and the imbalance between foreground and background pixel classes in medical images.

To improve feature utilization and address challenges, such as missing target regions and insufficient edge detail preservation, this study proposes a segmentation method that integrates path enhancement, residual attention, and zone-based chunking training.

The proposed method introduces a path enhancement structure consisting of a bottom-up path aggregation branch (PAB) and a multilevel fusion and complementary enhancement branch (FEB). The PAB aims to improve the transmission of semantic and positional information, while the FEB provides a richer feature representation for mask prediction. Additionally, a residual block with directional frontier support and combinatorial attention is designed to focus on important content units and boundary features. To further refine segmentation, a chunking strategy is employed to enhance the extraction of fine-grained foreground details through localized processing.

The method was evaluated through extensive ablation experiments, demonstrating consistent performance across multiple trials. When applied to lung nodule segmentation in computed tomography (CT) images, the method showed a reduction in mis-segmented regions. The experimental results suggest that the proposed approach can improve segmentation accuracy and stability compared to baseline methods.

Overall, the proposed method shows promise for medical image segmentation tasks, particularly in applications requiring precise delineation of complex structures.