U-Net-based medical image segmentation algorithm

Abstract

Before the deep learning algorithm, there are many traditional image segmentation methods. However, with the rapid development of artificial intelligence technology, the requirements for the accuracy and efficiency of image segmentation are getting higher and higher, so the image segmentation algorithm based on deep learning arises as the times require. But it is worth noting that these algorithms are natural image processing, and medical image format diversification, the difference of pixel value range, the presence of noise and artifacts, and so on, using the general image segmentation algorithms cannot meet medical demand scenarios medical image segmentation. On the basis of U-Net model, this paper improves its sub-modules, and proposes four sub-module structures: Path direct connection, Dropout direct connection, Conv direct connection, and Constant scale. Dataset uses Skin lesions of melanoma, and accuracy, sensitivity, specificity, precision, F-Measure, IOU, Dice coefficient and comprehensive score were selected as model performance evaluation indexes. After model training, validation and testing, the conclusion was drawn: In the segmentation task of melanoma skin disease image, the performance of Constant scale model was the best among all experimental models.