Medical Image Segmentation Based On Fully Convolutional Neural Network

In recent years,deep neural networks developed rapidly in the image field,which has made high-precision and automated segmentation of medical images possible,and this is also a great significance for doctors in rapid diagnosis,pathological analysis,and the development of wisdom medical system.As one of the cancers with the highest fatality rate,the pancreas has attracted widespread attention in the medical community.Due to the small size,changeable shape of the pancreas and the close fit of the surrounding tissues and organs,the characteristics of early lesions are not easy to be discovered in time,leading to the fact that the detection is often in the late stage.Therefore,the development of a high-precision,automated pancreas segmentation method is of great significance for clinical medical diagnosis.As one of the organs that have the most contact with the outside air,the lung is very susceptible to respiratory diseases.Among them,the new coronavirus attacks the human lungs by spreading in the air,making it difficult for the infected person to breathe and slowly invading other important tissues and organs of the body.It is even life-threatening,so the detection of lung diseases is very important for human health.At present,the most commonly used lung detection technology is CXR(Chest X-ray).Doctors conduct diagnostic analysis by analyzing lung characteristics in CXR.The difficulty of this task is that there are more dense abnormal substances around the lungs.These abnormal substances may be caused by diseases such as tuberculosis or pneumonia.Due to the high opacity,accurately segmenting the lung area is an important task.Challenging task.Aiming at the above-mentioned problems,this thesis conducts in-depth research on the basis of fully convolutional neural networks.The main work includes the following two aspects:1.Aiming at the problem that the traditional automatic segmentation method cannot achieve the ideal segmentation accuracy due to the small size and changeable shape of the pancreas,this thesis uses the idea of high-level semantic features to guide low-level features,and proposes a U-shaped convolution based on double decoding The singlestage pancreatic segmentation model of the neural network.The model consists of an encoder and two decoders.The two decoders use features of different coding depths to effectively combine low-level spatial information with high-level semantic information,and strengthen the efficiency of the segmentation network on feature information.It can be used to achieve high-precision segmentation of CT(Computed Tomography)slices that are not cropped and not reduced in resolution.Experimental results show that the method in this thesis can achieve better segmentation performance under full-scale input.The above segmentation effect is better than the existing single-stage pancreatic segmentation method.2.Aiming at the problem of low segmentation integrity caused by overlapping abnormal diseased tissues in the lungs,this thesis proposes a lung segmentation method based on a dual-encoded U-shaped convolutional neural network,which incorporates new encoding branches into multiple receptive fields Module to increase the network’s extraction of global feature information.In addition,in order to reduce the influence of irrelevant tissue regions on the segmentation results,the attention module is added to the jump connection between the encoder and the decoder of the segmentation network to improve the segmentation effect of the lung area.