Detection And Matching Of Calcification In Multiview Mammograms

Breast Cancer is the most common type of cancer in women.The appearance of breast calcification is an important signal of early breast cancer.Early Detection,early diagnosis and early treatment can reduce the mortality rate of breast cancer to some extent.At present,the main means of early screening for breast cancer is mammography.In mammograms,calcifications usually appear as bright spots,however,calcifications are usually very small,as well as with the lower resolution of the surrounding environment,the detection of calcifications especially the micro-calcifications is still difficult for radiologists.While an experienced radiologist will always observe the same lesion in multiple views to make a diagnosis,it can be a challenging task for an inexperienced radiologist to understand the relationship between the lesions.Therefore,if the computer can assist to detect the calcification and give the relationships between the lesions from the two views,it will greatly reduce the workload of doctors.Firstly,the existing detection methods of calcification are often based on single view,which has the problem of insufficient information.Secondly,the existing detection methods based on single view always extract manual features,which has a large number of false positives and the result were easily affected by breast density.Another method will use the existing detection framework to detect calcifications,while the calcifications are very small in the image,the number of positive anchors will be very small,so as to the sample imbalance problem will exist.In order to solve these problems,this thesis proposes a calcification detection and matching network based on multi-view mammography images,which aims to detect calcification lesions from different views and provide the matching relationship of them,so as to get the detection result based on multi-view images.Specifically,we studies from the following aspects:1)On the problem of calcification detection,while to extract manual feature is always very complex,this thesis adopts the segmentation method based on depth learning,which has good performance in the field of medical image at present,to extract the features of small lesions automatically.At the same time,the information of up-sampling and down-sampling is combined in the process of feature extraction to get more accurate segmentation results.Aiming at the problem of insufficient information in single view,the detection results of calcifications from different view were matched to reduce the false positive detection,and finally get the result based on multi-view image.The experimental results show that the accuracy of calcification detection can be improved effectively by using multi-view image information.2)On the problem of calcification matching,aiming at the problem that calcification always accounts for a small proportion in the image,so as to the network is difficult to extract the features of calcification effectively in the matching process,a matching network with attention mechanism is proposed.The attention graph is constructed by calculating the cosine similarity between the feature maps from two input images,then the attention graph and the feature are weighted average to get the features after the attention is drawn,then send it to the full connection layer for matching.The experimental results show that the attention mechanism can improve the matching accuracy.3)Finally,about the data problem,there are few public data sets can be used for breast mammography lesions matching at present,all the existing work will select the case with only one lesion for matching,this thesis collected and assisted to label a number of cases with multiple lesions of the corresponding relationship,for follow-up lesion matching research is of great significance.