Pancreas Automatic And Brain Tumor Segmentation Based On Deep Convolutional Neural Networks And Level Set

The pancreas is a crucial abdominal organ in the human body,and the occurrence of pancreatic diseases,especially pancreatic cancer,is high.The overall survival time for brain tumor patients is roughly one to one and a half years.Clinical medicine specialists usually manually depict pancreatic borders on CT image slices and different subregions of brain tumors on MRI image slices and develop and evaluate treatment plans for patients based on the depicted borders.However,manual segmentation by imaging experts is time-consuming,complicated,and the segmentation results are easily affected by factors such as the expert’s professional skills and medical experience;therefore,how to efficiently and automatically segment medical images such as pancreas and brain tumors images has been a contentious issue of professionals and experts in recent years.Deep learning has developed particularly rapidly in recent years,and deep convolutional neural networks can be used to segment medical images such as the pancreas efficiently and achieve good segmentation results,but it is difficult to add anatomical prior knowledge to the segmentation model.To solve the above problems,this thesis combines the advantages of deep convolutional neural networks and traditional segmentation methods to form a hybrid model,which integrates the image information obtained from deep learning into the level set model to achieve high-precision segmentation of the pancreas.In the brain tumor segmentation task,U-Net lacks a suitable strategy to utilize the global scene features.The global pyramid pooling model can address this issue and achieve the feature maps with different sizes by different levels in the pyramid pooling models.These different levels of features are helpful to enhance the segmentation accuracy.This thesis proposes a neural network model composed of a U-Net with encoding and decoding structures with the residual module,and a feature pyramid module(DFP)based on dilated convolution,namely,DFP-Res UNet.The major contributions are provided as follows:(1)For the pancreas segmentation task,the dataset of the 2018 ISICDM Pancreas segmentation challenge is selected,and all images are processed in the “dicom” format.MATLAB software was used to pre-process the dataset,the preprocessing operations include grayscale normalization,center cropping,and finally obtained 2D images of size 256*256 containing only the region of interest of the pancreas and with reduced grayscale deviation.The preprocessed pancreatic data were data augmented using center crop,vertical flip,and horizontal flip to obtain more training and validation data,which were then fed into three neural network models for training and testing,the HED network which is sensitive to target boundaries,the Segnet network which retains pooled index information and the Res U-Net which incorporates a residual module.The neural networks were trained using a pixel-by-pixel weighted loss function to acquire coarse segmentation results for each network model,respectively,using the Tensor Flow and Tensor Layer frameworks.This algorithm can quickly segment the initial localization of the pancreas as a preparation for the next algorithm;(2)The principle of the level set algorithm is explained,and the coarse segmentation results of the neural network model are optimized using the level set algorithm.The method takes the results of the three neural network models to implement intersection processing as the algorithm’s initialization and then outputs the final pancreas segmentation results after level set optimization.Three evaluation metrics,DSC(Dice Similarity Coefficient),PPV(Positive Predictive Value),and TPR(True Positive Rate),which are commonly used in medical image segmentation tasks,are selected to evaluate the segmentation results in this thesis.The mean values of the evaluation metrics DSC,PPV,and TPR are 0.854,0.876,and 0.901,respectively,once the optimized segmentation results are compared to the expert segmentation results.While contrasted to other successful pancreatic segmentation methods,the algorithm developed in this thesis can be shown to be very effective in improving pancreatic segmentation accuracy in CT images,which can assist doctors in the diagnosis and treatment of pancreatic diseases.;(3)For the brain tumor segmentation task,this thesis proposes a 3D neural network model to automatically segment different sub-regions of brain tumors in multimodal MRI,which consists of a U-Net with a coding and decoding structure of the residual module and a feature pyramid module based on dilated convolution.Data from the 2018 Multimodal Brain Tumor Segmentation Challenge was selected as the experimental dataset for the relevant experiments,which contains285 cases in the training set and 66 cases in the test set,each containing four modalities of 3D data.Using the proposed method on the validation set of the 2018 Multimodal Brain Tumor Segmentation Challenge data,the mean Dice scores for different sub-regions were ET(Enhance Tumor)0.8431,WT(Whole Tumor)0.897,and TC(Tumor Core)0.9068,respectively.In addition,high sensitivity and specificity,as well as low Hausdorff distance,were obtained in this thesis.A comparison with other brain tumor segmentation algorithms also illustrates the superiority of this algorithm.