YOLO-SV: An Optimized YOLOv8 Network for Sublingual Vessels Extraction

Sublingual vessels are an important part of tongue diagnosis in traditional Chinese medicine (TCM). It is widely used in the diagnosis and treatment of digestive system diseases, tumors, diabetes, and other clinical diseases. Traditional expert diagnosis relies heavily on the expert's medical ability and clinical experience and the time cost is high, which cannot be adapted to the increasing practical needs of patients. This paper proposes a method based on an enhanced backbone network of YOLOv8, named YOLO-SV. YOLO-SV automates the extraction of the tongue back and sublingual vessels of patients and assists doctors in the efficient diagnosis of diseases. In the YOLO-SV method, the Large Separable Kernel Attention (LSKA) is added between the Concat and Conv layers of the SPPF module of the backbone to form the LSKA-SPPF module. Three hundred ninety-two images containing the tongue back and sublingual vessels of gastritis patients are constructed as a dataset to train, validate, and test the model. The combined experiments show that YOLO-SV has an accuracy of 96.6% and a mean average precision mean (mAP) of 0.958, for the extraction task of Sublingual vessels. It outperforms YOLOv8, Mask R-CNN, and U-Net, with a boost of 16.3% in accuracy and 0.518 in mAP. The work can accurately extract the sublingual collaterals, provide an objective basis for disease diagnosis, and promote the progress of efficient TCM tongue diagnosis.