Research On Breast Tumor Image Recognition Based On Deep Learning

Medical images are the main method of medical examination and mammography is the primary method for breast tumor and breast cancer screening.Because traditional mammography requires radiologists to perform manual procedures in hundreds of pictures.The screening will consume a lot of time and energy and even missed or misdiagnosed.As deep learning technology becomes more and more mature,more and more researchers begin to apply deep learning to the segmentation and recognition of medical images.The target detection algorithm is one of the most effective means for tumor detection in medical images.In this paper,a standard mammography data set for deep learning target detection is constructed,and an improved tumor detection model based on the YOLOv3 algorithm is studied and implemented.The main work of this paper is as follows:(1)For the breast tumor image data set that lacks the standard and can be used for deep learning training,according to the cooperation with the tumor hospital,nearly 2,000 mammography images are collected,and through data balance、data expansion、CLAHE algorithm preprocessing and partial data denoising,about 10,000 standard data sets can be constructed for training and testing.(2)Designed a target detection model based on YOLOv3 and applied it to breast tumor image training and detection experiments.Through continuous analysis and research of experimental results,the network model constructed was continuously optimized.The measurement loss method of IOU is replaced by the GIOU regression loss method to obtain a better positional relationship.Secondly,the original YOLOv3 residual module is designed as a small U-shaped structure.Each scale feature map corresponds to a U-shaped structure to improve the network.The ability to perceive local features eventually forms the model of YOLO-U for the detection of breast tumors with mammography images.(3)Based on the tumor detection model YOLO-U,the standard breast mammography data set is trained and tested,and the YOLO-U model is compared with the experimental results and performance indicators of other deep learning algorithms.The research results show that after replacing the IOU with GIOU and adding a small Ushaped network structure,the accuracy and recall rate indicators have increased significantly,which can greatly improve the inaccurate positioning of the YOLOV3 algorithm.This research can solve the tedious traditional manual recognition and maximize the use of medical images for accurate medical diagnosis.This research can help clinicians quickly extract lesion information,determine the type of breast tumors and has practical value.