SBDFNet: a deep learning-based multiclass segmentation method via boundary distance fields

Abstract With the development of artificial intelligence in the field of medical imaging, medical image segmentation technology has played an important role in clinical practice. Extracting the morphological and structural information of lesions from medical images can provide important basis for disease diagnosis and treatment decision-making, and effectively promote the quantitative evaluation and progression monitoring of clinical diseases. However, the complexity and diversity of lesions make it difficult to design and optimize segmentation algorithms, especially for lesions with complex boundaries. To solve this problem, this study proposed a segmentation network driven by the boundary distance field of the segmentation map, called SBDFNet. The network gives the boundary distance information of the segmentation map and combines the distance loss function to effectively improve the accuracy of lesion boundary segmentation. Secondly, a new residual multi-scale feature fusion and spatial aggregation module is designed to enhance the perception of image detail features and spatial features through multi-scale encoding. In addition, a side-out supervision mechanism is constructed to supervise the prediction information of the intermediate layer and assist in reconstructing accurate segmentation results. Finally, the morphological opening and closing operations are used to focus on the segmentation of connected and discrete regions, so that the model can fully perceive the global and local information of the lesion and improve the overall segmentation effect. The SBDFNet method is comprehensively evaluated on four datasets: a tumor and liver automatic segmentation, LiTs, BraTs2019, and Spine, verifying its reliability and effectiveness. The proposed approach exhibits superior performance, reaching an intersection over union of 93.757% and a dice similarity coefficient of 96.7635%, which highlights its effectiveness in the segmentation task.