Method and device for fully convolutional single-stage breast image lesion detection based on multiple images

Abstract

Aiming at the fact that the existing breast lesion detection algorithm cannot well combine the information of bilateral breasts, cannot simultaneously meet the needs of identification and detection of multiple diseases including lumps and calcifications, and has a general effect on asymmetric dense glands, the present invention proposes A fully convolutional single-stage mammography lesion detection method with fusion of multi-image information. Use a non-anchor-based method for lesion detection, extract features of different scales from the original image, fuse the features of different scales, and fuse the information of different images, and finally directly predict whether a point on the feature map corresponds to a lesion and the specific location of the lesion.

Description

technical field [0001] The present invention relates to the field of image processing, in particular to a processing method for detecting lesion regions in mammary gland images. Background technique [0002] Breast cancer is the malignant tumor with the highest incidence rate in women. There are more than 270,000 new breast cancer cases in my country every year, and the incidence of breast cancer is increasing year by year, seriously threatening women's health. Early diagnosis of breast cancer is very important, early accurate diagnosis can increase the 5-year survival rate of breast cancer patients from 25% to 99%. Screening mammography is considered the test of choice for breast cancer screening. At present, mammography relies on subjective diagnosis, and the overall accuracy rate is not high enough and is limited by the level of evaluators. Compared with the personal experience judgment of a medical expert, the artificial intelligence recognition algorithm may identify ...

AI Technical Summary

Problems solved by technology

[0005] 1. The existing anchor-free detection methods generally only use the classification results located inside the object and the regression results of the rectangular frame of the object when identifying and regressing lesions, which will cause some low-quality prediction frames (and real The prediction frame with a large difference in the position of the rectangular frame) gets a higher classification score, while some high-quality prediction frames get a lower classification score, which further affects the subsequent non-maximum suppression process (NMS process, used to remove the non-maximum The result of the optimal prediction frame) makes the detection frame inaccurate

Moreover, the anchor-free method commonly used in the prior art is only aimed at mass lesions, and fails to cover all types and shapes of lesions, especially lesions with uneven shape distribution (such as calcification), thus resulting in missed detection

[0006] 2. The existing detection methods fail to make better use of the feature information on the left and right sides when using bilateral breast information for breast cancer diagnosis, ignoring the ability of the network to automatically mine and extract bilateral features

[0007] 3. Most of the existing lesion detection models do not make good use of the dual-view information of the same breast for breast cancer diagnosis, relying only on single-view images for detection, often misdiagnose normal glands as lesions due to gland overlap, or miss diagnosis inconspicuous lesions;

[0008] 4. Most of the existing lesion detection models are based on a two-stage detection method. The two-stage method is slow and takes a long time to complete the diagnosis

[0009] 5. The existing lesion detection model is not effective for the detection of asymmetrical dense, because the detection of asymmetrical dense depends on the comparison of bilateral breasts at the same time, and it is easy to judge by combining the information of two perspectives of the same breast

Images