A real-time segmentation and recognition method for dynamic ultrasound breast nodules based on deep learning

Abstract

The invention relates to the technical field of medical image processing, and aims to provide a method for real-time segmentation and recognition of dynamic ultrasound breast nodules based on deep learning. Including: collecting ultrasound breast images with nodules, video and case data with surgical pathological results to construct a data set, constructing a static image nodule segmentation network, training a static image nodule segmentation model on the original image; using the LSTM layer to predict intermediate frames Nodule probability, construct video dynamic segmentation network, train dynamic segmentation model; use basic network to construct benign and malignant recognition network structure, train benign and malignant recognition model; output nodule location information in real time, use benign and malignant recognition model to identify benign and malignant nodules in each frame , the end of the check outputs the number of nodules and the combined benign and malignant probabilities. The invention can avoid the information incompleteness of a single image, reduce false detection, reduce missing small nodules, and improve the accuracy rate of benign and malignant nodules.

Description

technical field [0001] The invention relates to the technical field of medical image processing, in particular to a method for real-time segmentation of dynamic ultrasound breast nodules based on deep learning and identification of benign and malignant nodules. Background technique [0002] Breast cancer has a high incidence worldwide, and early detection of breast cancer can greatly improve the survival rate and quality of life, and improve prognosis. The non-invasive auxiliary examination methods commonly used in breast cancer screening mainly include mammography, ultrasound and magnetic resonance. Among the three examination methods, mammography is cheap and mainly uses X-rays, which has the most advantage in detecting micro-calcifications, and can detect asymptomatic or unpalpable tumors. The diagnostic efficiency is even higher than that of magnetic resonance imaging. The advantages of ultrasound are that it has no radioactivity, is convenient to check, and is inexpens...

AI Technical Summary

Problems solved by technology

From the practice of video segmentation tasks, it can be seen that LSTM achieves better results than traditional time series models, but the basic network for practical application of LSTM is very shallow. If there are more training samples, a deeper network needs to be designed to improve model performance.

On the other hand, due to the black-box nature of deep learning, a large number of experiments are required to adjust parameters, such as the number of LSTM network layers, sequence length, training batch size, etc. These parameters may have a great impact on the results. Lack of effective theoretical guidance

Images