Application Of Deep Learning In Medical Image Segmentation

Medical image segmentation is the key to computer-aided diagnosis using medical images.In recent years,deep learning has achieved remarkable results in the field of computer vision,and Convolutional Neural Network(CNN)has also achieved a major breakthrough in image segmentation.Iron deposits in nucleus such as Red Nucleus(RN),Substantia Nigra(SN)and Subthalamic Nucleus(STN)are considered to be closely related to the occurrence and development of Parkinson's syndrome The accurate segmentation of these nuclei is a prerequisite for quantitative research on iron deposition in nuclei,and is of great significance.However,due to the small size of the nuclei and the poor contrast with the surrounding tissue,accurate segmentation is very difficult.The automatic segmentation of the pancreas area is a prerequisite for quantitative analysis of pancreatic cancer.Due to the large differences in volume and shape of the pancreas,and the blurred boundaries on the image,doctors often need to judge the boundaries based on anatomical knowledge and experience,and accurate segmentation is extremely difficult.Based on this,this paper attempts to use CNN to study the two medical image segmentation problems of deep gray matter nuclei segmentation in brain quantitative susceptibility mapping(QSM)and pancreas segmentation in abdominal CT images.In this paper,we used 2.5D Attention U-Net network with multiple inputs and multiple outputs to segment the RN,SN,and STN regions in high-resolution QSM images and transfer learning was used to address the problem of the small amount of training data.We studied the implementation details of transfer learning and found that best performance could be achieved when all the transferred parameters were fine-tuned.The final segmentation results were close to those of manual segmentation.In this paper,we proposed a 2.5D cascaded deep supervised U-Net combined with probability maps,namely CSNet,to segment the pancreas in CT images.We also trained 2D U-Net,2.5D U-Net,2.5D deep-supervised U-Net networks and compared their performance over the same dataset.DSC(Dice Similarity Coefficients),Precision,Recall,VOE(Volumetric Overlap Error)and RVD(Relative Volume Difference)were calculated to quantitatively evaluate the performance of the models.The experimental results proved that the segmentation accuracy of CSNet for pancreas was better than other traditional U-Net segmentation models.In this paper,deep learning is used to accurately segment deep nuclei in brain QSM images and pancreas in abdominal CT images.