A Study On Boundary-based Deep Learning Medical Image Segmentation Methods

As an essential part of computer-aided diagnosis,3D visualization,and medical images analysis,medical images segmentation has been playing a critical role in smart medical.Thanks to the rapid development of deep learning methods,medical image segmentation tasks have recently achieved substantial advances.The region-based segmentation method,which is prevalent in deep learning techniques,specifies the segmentation target in its entirety.Nevertheless,the boundarybased segmentation method that achieves segmentation by delineating contours and utilizing boundary details has gained increasing attention.Many deep learning methods leverage boundary information to produce segmentation with accurate boundaries.This dissertation proposes a series of boundary-based segmentation models from model and loss function perspectives and discusses how to accurately and effectively utilize boundary information.The contents and implications are illustrated as follows.Inspired by the traditional active contour,we propose a deep active surface model,which includes two stages,namely,surface initialization and surface evolution.We begin by using a convolutional neural network to produce an initial segmentation result and then estimate the first surface cloud points based on that result.The second step involves the construction of a cloud point network to predict the offset of the initial surface points,which evolves the initial surface to the desired position and enhances the accuracy of the initial segmentation findings.We examine the deep active surface model using two spleen datasets and find superiority in both accuracy and speed compared to other boundary-based methods.We propose a boundary refinement model that utilizes a two-stage approach to optimize the segmentation findings of boundary areas.The refinement stage aims to refine the boundary area of the coarse segmentation obtained in the first stage.To better control the process of boundary learning,we introduce the signed distance function(level-sets)in the loss function,and the boundary can be described by the zero level-set naturally.By utilizing the signed distance function,the network can extract geometric features of the boundary,and guide the network to learn the boundaries better.Additionally,the refinement step may be utilized as post-processing to optimize boundary segmentation in conjunction with other established approaches.We evaluate our model on multiple organs,and the experiments demonstrate the effectiveness and e ciency of the coarse-to-fine boundary refinement model.We propose a multi-level structural loss function,which measures the similarity between the predictions and the labels from multi-levels,and focuses on the boundary areas.We summarize the region-based and the boundary-based loss functions,then propose a novel pixel-wise loss function to deal with the unsatisfactory boundary.The regional loss term depicts the segmentation object as a whole,the boundary loss term describes the contour information,and the pixel-wise loss term supplies the details on pixels.The three loss terms are complementary,which helps to extract the key features.We test the loss function on multiple datasets,and the experiments show that our loss function not only improves the accuracy of segmentation,but also accelerates convergence.We further propose a multi-level structural network to make the best of regional,boundary,and pixel-wise features.The network architecture contains three branches,which are used to extract the features of the region,boundary,and uncertain points(pixels).Then a multi-level saliency guidance module is used to fuse the features of the three branches,where the features of boundary and the uncertain points operate as supplements to the regional branch to produce segmentations with accurate boundaries.We compare our model to other methods on multiple pancreas datasets and spleen datasets,and our model achieves a superior segmentation.