Classification And Detection Of Cancerous Region In Mammographic Images Using Deep Computing

In the process of breast cancer diagnosis, analyzing the digital mammography images of the lesion area is one of the most common used breast cancer examination. Mammography images are usually collected by penetrating the breast soft tissue using soft rays. However,the diagnostic results always tend to have a strong subjectivity, owing to the doctor’ s time consuming diagnostic method and subjective diagnosis, which may have a detrimental influence to the patient. The computer aided diagnosis system can provide diagnostic results objectively and accurately, which can provide auxiliary analysis accurately and quantitatively for patients to have a more suitable treatment option, and reduce the doctor’s heavily diagnosis process,which are of great significance. Mammography images are widely used for the reason of simple,fast,painless and cheap. The lesion area in Mammography images have various types, complex structure, and no consistency, and normal areas and lesion areas have high similarity. Therefore, the classification and detection of lesion areas is a very challenging job.For the above questions, this thesis presents a deep convolution neural network based method of automatic classification and detection of lesion areas in mammography images. Firstly, the sliding window is used to solve the problem of size inconsistency between the lesion areas and the normal regions in the mammography images. Then we use deep convolution neural network to train classifier based on different deep feature, and compare it with the feature generated by traditional ways and the fusion of traditional feature and deep feature. The experimental results show that considering different features,the fully connected layer, FC6, which is the sixth layer of the deep convolution neural network achieves the highest value of F1, 0.8902, which describes the robustness of the classifier.Based on the results of classification, the automatic detection of lesion areas in mammography images is further constructed. The lesion areas marked in the data are updated to the smallest rectangle which include the lesion areas.We reconstruct data sets, and train the convolution neural networks. The sub part of mammogram image which is achieved by sliding windows is taken as the input of the trained deep convolution neural network, and then judge whether the sub-image is cancerous. Then we can get the suspected area of cancer by updating the sub-images to rectangle. Reconstructing the data set according to the lesion areas and the areas where the cancer may exist, the new convolution neural network is trained by the above method to further judgement of whether the sub-image of the region is a cancer sub-image, if the confidence value is below 0.95. Finally, the validity of the method is verified by quantitative analysis, the method we proposed can be verified by quantitative analysis. The final detection using our approach achieves the mean positive of 0.89 and there are 3.61 false positive sub-image.