Segmentation Of Pancreas And Lesions From Abdominal CT Images Based On Deep Convolutional Neural Network

Accurate segmentation of pancreas and lesions in computed tomography(CT)of abdomen images is an important step in computer-aided systems for tumor analysis,clinical diagnosis,surgical resection planning,and preoperative planning for tumor ablation.With the development of deep learning technology and its outstanding power of learning,deep convolutional neural based segmentation network has developed to be the most prevailing algorithm applied to medical image segmentation.However,traditional segmentation methods based on deep learning are still challenged to obtain meticulous segmentation results,the main reason is that the deep learning methods are always affected by a large area of background and the large anatomical variations of shape and position of the pancreas in abdominal CT images.Furthermore,segmentation methods based on deep learning often require large-scale pixel/voxel-level annotated datasets for training and testing,which is difficult to achieve in medical image segmentation tasks.Therefore,this thesis is devoted to the study of some key problems in pancreas segmentation based on deep convolutional neural network.This thesis mainly includes the following three works:(1)Automatic segmentation of pancreas based on coarse localization and ensemble learning.The pancreas usually occupies a small proportion in abdominal CT images and has low contrast with surrounding tissues.Meanwhile,the anatomical shape and position vary greatly from patient to patient.To address these issues,we propose a coarse-to-fine two-stage segmentation framework.In the stage of pancreas coarse localization,we adopt a bottom-up strategy considering the influence of larger background region for the CNN.The local image patches based on super-pixels in the three views are classified by the ResNet to obtain the coarse labels of the whole image.In the fine segmentation stage,we use the pancreas region obtained from the previous stage as the input of the segmentation network,and concatenate the image patches from its two adjacent slices to introduce 3D information to help the network learning.For problems such as category imbalance,we design multiple loss functions with different aims for training and obtain multiple segmentation networks with different focuses.Finally,we get the final segmentation result with an averaging-based ensemble algorithm which ensures the segmentation accuracy under various bad situations.(2)A semi-supervised automatic segmentation algorithm of pancreas based on graph-enhanced network and uncertainty guidance.Pancreas segmentation and medical image segmentation are often faced with a little bit of labeled data at the pixel/voxel level.To address this problem,we propose a semi-supervised automatic segmentation algorithm for the pancreas based on a graph-enhanced network and uncertainty guidance.Based on the nn-UNet,we add a graph enhancement module based on graph convolutional neural network to model the relational between the deep semantic feature and spatial as well as gray information of voxels,which can help the CNN network to effectively utilize the global spatial distribution information and mine voxels’ labels from high-quality pseudo-label images with the global label propagation of graph structure.Considering the possibility of mis-segmentation in the generated pseudo-labels,we introduce a semi-supervised learning pipeline built with an iterative uncertainty-guided pseudo-label fine-tuning strategy.Through the evaluation and selection of pseudo-label quality,the network can learn more semantic information from high-quality pseudo-labels.(3)Pancreas and lesion segmentation algorithm based on deep feature discriminative learning.By analyzing the relationship between deep features and segmentation accuracy in the pancreas and lesion segmentation tasks,we find that the more aggregated deep features of the same category,the higher the segmentation accuracy.Motivated by this,we propose a deep feature discriminative loss function.From the view of deep features,this loss encourages the aggregation of the pixels’ deep features of the same category,and keeps the aggregation centers of the deep features of different categories away from each other.By restricting feature discrimination,the influence of class imbalance in label constraints on the optimization process can be reduced.Experiments show that this loss function can be applied to other segmentation networks.