Research On CT Image Segmentation Method Of Liver And Tumor Based On Deep Active Contour Models

Liver cancer is one of the liver diseases with high morbidity and mortality.Liver cancer treatments require accurate diagnosis and planning.Computed tomography（CT）scan is the most commonly used imaging techniques for liver cancer diagnosis since it gives accurate anatomical information about the abdominal organs in the human body.Manual segmentation of liver organs and tumors from CT scans is tedious and timeconsuming.Currently,computer-aided diagnosis has been widely used to segment liver and liver tumors.However,large variations exist in the location,shape,and size of organs among individuals.Low contrast and blurry edges are the main characteristics of CT images,which make automated liver delineation a chalenging task.Tumor segmentation adds more challenge due to the small observable changes between tumors and healthy tissues,especially at their borders.The active contour models are classical methods used for boundary detection and target segmentation,and the segmentation is achieved by deforming the initial contour of a 2D image or the initial surface of a 3D image toward the target boundary.However,the method still relies on manual initialization and the location of the initialized contour needs to be in the vicinity of the target region.Recently,the development of deep fully convolutional neural networks enhanced the performance of the semantic segmentation and leads to outperforming other competitors in the field of medical imaging.However,constant convolution and pooling operations in the network can lead to feature loss.Moreover,the network usual y performs pixel-level prediction independently,and it is difficult to take into account the consistency of local features and the smoothness of edges.Despite the high accuracy achieved by FCNs in segmenting organs,neural networks tend to ignore the semantic information of the boundaries for irregular lesions.Several studies have integrated the advantages of classical energy-based active contour models and modern learning-based CNNs in various segmentation tasks due to their complementary characteristics.These methods can be classified into three categories: first,using active contour models as post-processing for CNNs;second,incorporating ACM into deep learning methods in an end-to-end fashion;third,reformulating the energy functional as loss functions.In abdominal CT images,the liver is relatively large and the boundary is regular,the ACM post-processing method can improve the segmentation accuracy and optimize the segmentation boundary.However,liver tumors are small in size and irregular in shape.Only post-processing of the active contour model cannot improve the segmentation accuracy.Therefore,an end-to-end method is proposed to guide the model to locate the tumor region and achieve accurate segmentation.To preserve the spatial context information of medical images,the structures proposed in this paper are all based on 3D models.The specific research contents are as follows:（1）Due to low contrast and blurry edges of the liver in CT images,a two-stage segmentation method based on the V-net and Chan-Vese models is proposed,cal ed3 DAC.This method uses the multi-branch V-net model in the first stage to obtain the segmentation probability map and the initial level set function（signed distance functions）used for prediction.The initially signed distance map not only serves as an initialized level set function for the CV model but also guides the V-net model to determine the location region of the liver.In the second stage,to eliminate the discontinuity of the ful y convolutional neural network prediction,the CV model was used as a post-processing step and its evolution was extended from the 2D contour to the 3D surface,and a smoother 3D visualization result was obtained by directly evolving the whole liver surface.The advantages of the proposed method for liver segmentation were validated on the Abdomen CT-1K dataset.Compared with the DALS method,the Dice coefficient and the Jaccard coefficient are improved by 4.29% and4.58%,and the ASD and 95 HD indicators are improved by 5.96 mm and 10.04 mm,which proves the effectiveness and robustness of the proposed method.（2）An end-to-end trainable 3D image segmentation framework,3DTAC,is proposed because liver tumors are smaller in size and have more complex boundaries in CT images,and cannot be better identified and segmented using only ACM as postprocessing.the framework consists of a convolutional neural network（CNN）and an ACM with learnable parameters.The object boundaries can be accurately detected using ACM global optimization and trained by end-to-end differencing.All four metrics of tumor segmentation on the Li TS dataset outperformed the DCAC method,demonstrating the effectiveness of the proposed method.