Research On Self-paced Learning And Deep Learning For Abdominal Image Segmentation And Enhancement

Medical image analysis is the most basic and import step in disease diagnosis and treatment.With the continuous development of computer technology,using image processing methods and machine learning is becoming a new direction of medical image analysis.However,the small amount of medical image and the great difference between human beings make more difficult for medical image analysis using traditional machine learning methods and deep learning methods.Pancreatic cancer is one of the most malignant cancers.Accurate segmentation of patients' pancreatic images is helpful for diagnosis and treatment.In this paper,traditional machine learning methods and deep learning methods for segmentation of pancreatic images with different modality and few training samples are proposed.In addition,with the increasing attention to radiation injury caused by medical imaging equipment,degraded medical image is becoming more and more widely used.In order to further promote application of degraded medical image in analysis like segmentation and registration,this paper also concentrated on enhancement of degraded medical image using deep learning.Aiming at the above problems,the following three tasks have been completed:1.To solve the problem that pancreas in 3-dimensional MRI cannot be accurately segmented,which is small and adhesive with surrounding tissues,a sequential segmentation model using level set and self-paced support vector machine learning(SPSVM)is proposed.Firstly,the approximate contour of the pancreas in the first image of MRI sequence is drawn by human-computer interaction,and the rough segmentation results of the sequence are automatically evolved by level set.Then,the rough segmentation results are further segmented by the trained SPSVM.Finally,morphological methods are used as postprocessing to get final results.In the process of training SPSVM,self-paced learning is used to optimize training process of SVM,which learns simple samples at early steps and then gradually involves more complex samples in training process.Besides,we construct the input of SPSVM by combing gray-level co-occurrence matrix,gray level-gradient cooccurrence matrix and its three dimensional expression.The experimental results show that the proposed method improves the accuracy of segmentation and have effects on removing adhesion.2.In order to deal with the problem of pancreas segmentation in PET/MRI,whose MRI images are two dimensional and have low contrast and poor quality,a segmentation model based on transfer learning and multi-scale convolution is proposed.Firstly,a U-Net model who has encoder-decoder structure is trained using PET images.Secondly,a MRI segmentation network based on U-Net,which has dual encoders and single decoder is proposed.In MRI segmentation network,one encoder transfers from PET U-Net to help MRI segmentation network pay more attention on region of interest and multi scale convolution module is used to improve feature expression for another encoder.In the decoder part,the feature maps are performed by interpolation upsampling and deconvolution to output the MRI segmentation results.The experimental results show that the idea of transfer learning can better locate the pancreas,multi-scale convolution and fusion of two upsampling methods improve the segmentation accuracy.3.Aiming at the problem of severe noises and low resolution in MVCT images,an enhancement model based on generative adversarial network with reconstruction loss is proposed.Learning the mapping relationship between MVCT which has low quality and KVCT whose quality and resolution is high is the core of this method.A generator with double input of MVCT and KVCT is constructed.Cycle-consistency loss is used to solve the problem of unpaired input data,and reconstruction loss is introduced in the part of reconstruction in generator to improve quality of generated image,in addition,gradient term is used to keep the contour of the generated image corresponded with MVCT.Experimental results on real clinical MVCT images show that our method gets excellent visual effects and high evaluation criteria.