Surgical Region of Interest Detection from Laparoscopic Images Using Boundary Fuzziness

During laparoscopic cholecystectomy (LC), automatic Calot’s triangle region of interest (CTROI) recognition can assist surgeons enhance the safety and efficiency of surgical procedures. Due to the absence of specific color, texture, or shape features of CTROI, automatic recognition is extremely challenging. Most of the state-of-the-art methods in medical image recognition are based on end-to-end deep learning approaches, which require vast accurately annotated medical images to train. However, the complexity of medical anatomical structures often leads to ambiguity among annotators, resulting in label variability. In this paper, we propose a step-wise relaxation method that processes the original labels to describe the variability of annotations through intermediate states. Furthermore, considering the relationship between the locations of the gallbladder and CTROI, we optimize the SegFormer network structure. The improve structure combines the segmented gallbladder with the feature map, providing an attention mechanism for CTROI detection in images. Cross-validation experiments demonstrate that our methods hold advantages on dice similarity score and intersection over union over SegFormer.