Endoscopic Segmentation of Kidney Stone based on Transfer Learning

Endoscopic stone segmentation is of great significance in the comprehensive diagnosis and surgical planning of ureteroscopic lithotripsy. Due to the quality problems of endoscopic imaging, such as artifact, highlight, reflection, contrast imbalance and blur, it is a great challenge to segment kidney stone fragments of different shapes. In this study, the improved U-Net model was used to extract the kidney stone fragment area at the pixel level, and its contour information could be accurately obtained. The VGG16 network with strong portability is used as the encoder of U-Net model to extract the semantic information of multiple feature layers, and the up-sampling is realized by using transposal convolution to gradually restore the segmentation details. In the experiment, DCE Loss combining Dice Loss and Cross Entropy Loss is adopted as the Loss function of the model. The experimental results show that the improved U-Net model has higher accuracy for stone segmentation from endoscopic images, with the MPA, MIoU and F1 score of 96.44%, 97.62% and 97.03% respectively. The F1 score of the modified model is 2.25% higher than that of Deeplabv3 + model, and 14.24% higher than that of standard U-Net model.