Mammography Mass Segmentation Model Research Based On FCN

At present,breast cancer is one of the most common malignant diseases of women worldwide.The key to reducing breast cancer mortality is early diagnosis and treatment.Breast mass is the main symptom of breast cancer,and precise mass segmentation is an important part of early detection of breast cancer,which directly determines the subsequent feature extraction and classification.However,the sizes and shapes of the masses are different,the edges are blurred and diverse,and it is difficult for medical imaging specialists to determine the location of the mass in the complex breast structure correctly.It results in misdiagnosis or missed diagnosis.Breast mass segmentation in Mammograms based on traditional digital image processing method often requires manual participation.And these methods are cumbersome and time-consuming.The deep learning method can automatically learn image features and establish a nonlinear complex model from input to output.It has been widely used.In order to improve breast mass segmentation accuracy,this paper studied the deep learning models of breast mass segmentation.The main work and innovations of the thesis are as follows:1.Aiming at the problems caused by breast masses with the blurring and the complex gradient boundary,the encoder-decoder breast mass segmentation model based on FCN is proposed.The idea of the model is to make full use of the high-resolution information extracted by the shallow layer of network.Add the information to the up-sampling network by skip connections to complement the missing boundary information during the up-sampling process.It can improve the prediction accuracy of the mass edge.2.The use of the encoder-decoder segmentation model may cause the problems of excessive use of computational resources,redundant model parameters,insufficient extraction and utilization of high-level features and multi-scale information,and poor results of complex breast tissue images segmentation.In order to solve these problems,an automatic mass segmentation model integrated withattention mechanisms and dense connections based on FCN is proposed.The idea of the model is to use densely-connected convolutional neural network as feature extractor to realize feature reuse.Dense connections can alleviate the gradient disappearance caused by network deepening.The introduction of attention gates make gradients originating from background regions are down weighted during the backward pass.The update of network parameters mostly depends on the feature response of the mass area.The network focuses on learning the features of the masses.It does not need to introduce additional object detection networks that will make the model parameters redundant,and reduce the influence of complex breast tissue on the segmentation results.It can achieve a more accurate mass segmentation.3.To verify the validity of the proposed segmentation model,a comparative experiment is performed on the mammograms in DDSM database and the segmentation performance is evaluated by F1-score,sensitivity,specificity,overall accuracy,ROC curve and AUC.The experimental results show that the proposed encoder-decoder segmentation model obtains ideal segmentation results when the mammograms with not complex breast tissue,but the segmentation results are not so good on mammograms with complex breast tissue.The proposed segmentation model integrated with attention gates and dense connections reduces negative effects of complex breast tissue and improves the segmentation accuracy without introducing redundant parameters and occurrence of gradient disappearance.