Breast ultrasound image tumor segmentation method

Abstract

The invention discloses a breast ultrasound image tumor segmentation method. The method comprises the steps of breast ultrasound image data preprocessing, deep neural network model construction, loss function definition, model training and result generation. In data preprocessing, a mode of firstly filling a mirror image and then cutting is used, so that the form of a breast tumor is not changed, and a breast ultrasonic image meeting the size requirement can be obtained. In the step of constructing the deep neural network model, the design mode of the UNet model is followed on the whole. According to the method, ResNet18 is used as an encoder of the whole network, so that the method has higher feature extraction capability, and higher precision can be obtained; meanwhile, deep supervision technology is used in a model decoder part to supervise learning of each layer, and a channel-by-channel weighting module of SENet is added; thus, the problems of wrong segmentation and discontinuous segmentation boundaries can be eliminated, and the tumor boundaries can be accurately captured.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and is an effective new method for tumor segmentation in breast ultrasound images. Background technique [0002] Deep learning has become a hot spot in artificial intelligence. Deep learning technology is a machine learning technology that has developed rapidly recently. It uses deep artificial neural networks as a model, and uses powerful hardware, massive data, and optimization algorithms to achieve high-performance learning. Deep convolutional neural network (Deep CNN) has achieved very good results in many computer vision tasks, such as target classification, target detection, target segmentation, etc., and has greatly surpassed traditional machine learning and artificial intelligence technology in many fields. [0003] Breast cancer is one of the most common causes of death in women. In the early stage, mammography and mammography can detect tumors, but it lacks a good ability to judge...

AI Technical Summary

Problems solved by technology

However, ultrasound breast images often have characteristics such as high noise, uneven gray scale, and complex and variable tumor shapes, which greatly increase the difficulty of ultrasound image segmentation and classification.

[0004] In recent years, some scholars have proposed different breast ultrasound image segmentation methods, such as active contour models, graph-based algorithms, etc. Such non-deep learning methods often require human participation in the segmentation process, with a certain degree of subjectivity; another example is UNet , Attention-UNet, UNet++ and other models, this type of deep learning method uses an end-to-end training method to avoid human participation, but it often has shortcomings such as mis-segmentation, under-segmentation, and discontinuous segmentation boundaries.

Images