Medical Image Segmentation Based On Fully Convolutional Neural Networks

The use of computer-aided diagnosis technology to analyze and process medical images can provide strong evidence for the diagnosis and treatment of diseases.Medical image segmentation is a key step to realize medical image processing.Traditional medical image segmentation relies on manually extracting features,which makes the segmentation task onerous and complex.In recent years,with the development of deep learning,people can use the excellent feature learning ability and end-to-end processing mode of fully convolutional neural networks for medical image segmentation.In this paper,mass segmentation in mammography images and retinal blood vessel segmentation in fundus images are the research contents.The specific details are as follows:1.U-Net,a commonly used fully convolutional neural network for breast mass segmentation,has the following problems: 1)The limitation of skip-connections makes U-Net unable to effectively extract the multi-scale features of the mass;2)The simple feature fusion method causes U-Net ignores interdependencies between channels.To alleviate the above problems,this paper proposes an improved U-Net on breast mass segmentation method,which includes an adaptive scale module(ASM)and a feature refinement module(FRM).The ASM is embedded in each level of skipconnections,which adaptively adjusts the receptive field according to the mass scale to improve the multi-scale feature extraction capability of the network.The FRM is introduced in the decoder,which is able to capture channel dependencies,allowing the network to enhance the feature representation of useful channels.The segmentation performance of our method is evaluated on the mammography databases of DDSM-BCRP and INbreast.The Dice coefficients obtained by our method are 91.41% and 93.55%,respectively.The experimental comparison shows that the segmentation performance of our method is better than U-Net and some advanced breast mass segmentation methods.2.Aiming at the category imbalance problem of samples,low contrast and complex anatomical environment of retinal blood vessels in fundus images,this paper proposes a retinal blood vessel segmentation method based on improved U-Net,which uses U-Net as the basic network and develops the category attention guidance(CAG)module at the deepest level of the encoder.The CAG module allows the network to explore the global context with the help of long-range dependencies by establishing an attention mechanism between pixels and class labels.A deep supervision strategy is introduced at the output layer of the encoder to supervise the encoder to learn high-level semantic details,thereby assisting the CAG module to capture more accurate dependencies.Compared with U-Net,the segmentation performance of our method is significantly improved on both DRIVE and CHASE?DB1 databases.Compared with the state-of-the-art methods,our method also reflects a superior level of retinal vessel segmentation.