Research On Mammography Mass Detection Based On YOLOv3

Breast cancer is the most common malignant tumor among Chinese women.Early detection with correct diagnosis is extremely important to increase the survival rate.In clinical practice,mammography is the most reliable tool widely used in the early screening of breast cancer.To correctly detect and diagnose breast cancer,radiologists must read a large number of mammograms every day.The continuing heavy workload inevitably cause visual and psychological fatigue,which leads to an adverse impact on the subjective judgment by radiologists.Therefore,it is necessary to utilize computer-aided detection and diagnosis techniques to provide second opinions to radiologists and assist them in making accurate diagnosis.With the advancement of modern technology,deep learning has played a key role in breast cancer diagnosis.However,breast cancer diagnosis based on deep learning remains to have the following limitations: 1)The contrast between the masses and the surrounding breast tissue in mammography images is low,so it is difficult to obtain better detection performance by using traditional methods directly.In addition,the lack of large-scale labeled mammography data is easy to cause the problem of over-fitting in the training process of the model;2)In mammography images,some masses occlude each other,which leads to false detection in the detection process;3)Due to the large proportion of small sized masses in mammography images,there will be missed detection in the detection process.For the above problems,this paper studies the mass detection technology in mammogram images by combining image processing technology and object detection algorithm,which is mainly divided into the following three parts:1)Preprocessing method of mammograms.To enhance the contrast between the mass and other breast tissue in mammogram.In this paper,the method of double hat transformation on morphology is used to enhance the information of the mass inside the structural elements while keeping the information of the tissue outside the structural elements unchanged,so as to achieve the effect of local enhancement.In order to solve the problem of lack of large-scale mammography data,this paper uses data enhancement methods such as Cutout,Cutmix and Mosaic to expand the mammography dataset,which enlarges the original data set by five times.The values of mAP@0.5 and F1 in the enhanced dataset are 4.18% and 3.11% higher than those in the original dataset,respectively,which improves the performance of breast mass detection.2)MSD-YOLOv3 mammograms mass detection system.In view of the issue that occluded masses are easy to be missed in breast cancer diagnosis based on deep learning,a MSD-YOLOv3 object detection network for breast mass detection is proposed in this paper.First of all,a bottom-up path is added into the feature fusion module,and the cascading and cross-layer connections are adopted to make full use of the underlying feature information to improve the recognition accuracy of small masses.Secondly,to filter out more accurate prediction bounding boxes and avoid missed detection of masses that occlude each other,the DIo U is introduced in Soft-NMS to suppress the redundant prediction bounding boxes.The experimental results demonstrate that the presented breast mass detection framework has a high accuracy in detecting occluded masses of breast masses.The mAP@0.5 reaches 95.48%,F1 reaches 94.83%,which outperforms the counterparts.3)RFD-YOLOv3 mammograms mass detection system.To solve the problem that small masses in mammography images are often missed diagnosis due to the difficult of feature extraction,this paper proposes an improved detection and classification method based on RFD-YOLOv3.Firstly,the method uses the recursive feature pyramid as the feature fusion module,which improves the recognition accuracy of the small masses.Secondly,in order to speed up the calculation and save the calculation cost,an improved depthwise separable convolution is proposed to replace the standard convolution in the backbone.Finally,the ablation experiment of the improved module is carried out and compared with other traditional object detection algorithms.The experimental results show that the mAP of the improved YOLOv3 breast mass detection method reaches 96.17%,which is 2.43% higher than that of the original YOLOv3 algorithm.The detection time of each image only needs 18 ms,which is shortened by 4ms.