| Breast cancer has become the largest cancer in the world,early screening is an important means to reduce the mortality of breast cancer.At present,convolutional neural network has become the mainstream research method in the field of object detection,and has been applied to the field of medical image detection.However,most of the existing target detection algorithms are to detect natural targets of random size,whose size is greatly different from that of the breast mass in mammography images.If the algorithm is directly used to detect the breast mass,the phenomenon of misdetection and missed detection is likely to occur,and the final detection accuracy is not ideal.In order to improve the detection accuracy of breast masses,based on the study of Mask R-CNN,this paper improved the feature fusion method of feature extraction part,improved the size of Anchor of RPN part,and modified the candidate box screening algorithm,and proposed a target detection algorithm,D-Mask R-CNN,which is suitable for detection of breast masses.Its main contents are as follows:(1)In the process of feature extraction,it is difficult for the predictive feature layer to obtain the feature information needed for target detection only from a certain feature layer.In the original Mask R-CNN,FPN was used for feature fusion,and the feature information contained in different feature layers was successfully fused into each predicted feature layer,but the utilization rate of low-level features was still very low.To solve this problem,this paper improved the internal structure of FPN.Referring to the Dense block structure in the Dense Net network,the Dense connection was used to replace the horizontal connection in the original FPN during feature fusion,thus strengthening the transmission of features and the utilization rate of low-level features,and thus improving the accuracy of target detection and positioning.(2)The original Mask R-CNN was designed to detect natural targets,such as cats,dogs,cars,etc.,and the size of the detected targets was significantly different from that of breast masses.Considering the actual size of the breast mass in the image,the anchor size of the RPN part was redesigned in this paper,and the size of the anchor in the original model was modified to(8~2,16~2,32~2,64~2,128~2),so as to improve the positioning accuracy of the breast mass.Secondly,the number of candidate boxes generated by RPN part is very large,including a large number of candidate boxes with high coincidence degree.The original Mask R-CNN used NMS to remove redundant candidate boxes.However,considering that a single image in some breast images contains multiple near-distance lumps,the algorithm may remove the correct predicted results incorrectly.To solve this problem,this paper used Soft-NMS to replace the NMS in the original model,so as to reduce the possibility that the use of NMS might lead to missed detection of breast masses.In this paper,the breast mass images in CBIS-DDSM dataset were preprocessed and labeled,and then all the experiments were conducted for training and testing.The experimental results showed that the m AP value of the improved D-Mask R-CNN for breast mass detection on the test samples reached 0.66,which was 0.05 higher than that of the original Mask R-CNN.It was proved that the D-Mask R-CNN method proposed in this paper has better detection accuracy and is more beneficial to the detection of breast masses than the original Mask R-CNN. |
PDF Full Text Request