Breast cancer MRI segmentation method based on hierarchical convolutional neural network

Abstract

The invention discloses a breast cancer MRI segmentation method based on a hierarchical convolutional neural network, and the method integrates a plurality of algorithms, and comprises the specific steps: training a full convolutional network FCN model, and obtaining the global and local structure information of an input image through employing a 3D U-Net architecture; training the other two FCN models by taking the breast mask as a guide, and respectively estimating a coarse segmentation result and a refined initial result; developing a detection model based on the mark points; detecting twomarkers for biopsy tumor selection and radiogenomics so that common problems in mammary gland DEC-MR image segmentation are well solved. The common problems include imbalance-like problems, confusionproblems and the like which are difficult to solve, and a learning framework is designed to segment mammary gland tumors from coarse to fine; the biopsy tumor in all the detected tumors is determinedby using the tumor position information and the marking information, and the biopsy tumor detection system is ingenious in automatic detection and design, high in accuracy and convenient to popularizeand use.

Description

Technical field [0001] The present invention relates to the technical field of medical technology, in particular to a breast cancer MRI segmentation method based on a hierarchical convolutional neural network. Background technique [0002] Breast cancer is the most common cancer among American women except skin cancer. China is not a country with a high incidence of breast cancer, but it should not be optimistic. In recent years, the rate of increase in the incidence of breast cancer in my country is 1 to 2% higher than that of countries with high incidence. According to the 2009 breast cancer incidence data released by the National Cancer Center and the Ministry of Health's Disease Prevention and Control Bureau in 2012, the incidence of breast cancer in the national cancer registration area ranks first among female malignant tumors, and the incidence of female breast cancer (crude rate) nationwide The total is 42.55 / 100,000, urban is 51.91 / 100,000, and rural is 23.12 / 100,000. ...

AI Technical Summary

Problems solved by technology

[0005] (1) Manual labeling: This is the most direct method. The radiologist manually labels the tumor area. This method is not only time-consuming but also error-prone.

[0006] (2) Atlas-based methods: Although various atlas-based methods (by image registration) have achieved satisfactory results, since breast tumors usually do not have fixed locations and regular shapes, such methods cannot be accurate breast tumor

[0012] Disadvantages: Regardless of the semi-automatic segmentation method, they are highly dependent on predefined slices or regions of the tumor (such as bounding boxes, etc.), which require labor-intensive annotations by radiologists

[0016] Disadvantages: Traditional learning-based methods often regard feature extraction and model training as two independent tasks, ignoring the possible heterogeneity between artificial features and subsequent learning models

However, the existing deep learning-based breast cancer DCE-MRI segmentation methods seldom consider the influence of confounding regions and the ubiquitous class imbalance problem in breast cancer DCE-MRI segmentation.

Images