2019 6th International Conference on Systems and Informatics (ICSAI) | 2019
Application of U-Net with Variable Fractional Order Gradient Descent Method in Rectal Tumor Segmentation
Abstract
U-Net is a classical unsupervised deep learning algorithm and a representative algorithm of full convolution neural network. From the perspective of algorithm structure, U-Net is an auto-encoder with convolution and deconvolution operations. Because in the optimization part of the algorithm, the traditional gradient descent method is easy to fall into the trap of local minimum, and its global convergence is weak. However, the fractional gradient can retain the gradient information better and has strong global convergence. The main purpose of this paper is to introduce variable fractional order gradient descent method into the optimization part of gradient descent method, and use this algorithm to solve rectal tumor segmentation.