Pulmonary Nodules Segmentation Based on Multilevel Fusion Path and Data Augmentation

Medical image segmentation is pivotal in computer-aided diagnosis. Although U-Net has been widely used in this field, its performance is limited by challenges such as the incomplete transmission of feature information in medical images, the lack of labeled data, and the imbalance between foreground and background pixel categories. To improve feature utilization and address issues of missing details, a segmentation method that incorporates multilevel fusion block, residual mixed attention, and data augmentation is proposed. The method includes a multilevel fusion path with a bottom-up path aggregation (BPA) branch and a complementary enhancement (CE) branch. The BPA branch optimizes the transmission of semantic and positional information, while the CE branch provides a richer feature basis for mask prediction. Furthermore, a residual block with directional edge support and mixed attention to focus on vital content units and boundary features is designed. In addition, the data augmentation strategy with local components improves the extraction of fine-grained foreground areas and enhances the diversity of data. Segmentation of pulmonary nodules in computed tomography (CT) images demonstrates a reduction in mis-segmented regions. The experimental results indicate improved segmentation performance and stability with the method. Overall, this approach is promising for pulmonary image segmentation applications.