Research On Medical Lung Image Segmentation Method

Medical image segmentation is the process of segmenting the region of interest in the medical image accurately.It is the basis of further processing and analysis of medical images by a computer-aided diagnosis system.The segmentation result can be regarded as a reference for subsequent disease diagnosis,treatment plan design,and treatment effect evaluation.Recently,both Computed Tomography(CT)and Chest X-Ray(CXR)techniques have become research hot topics at home and abroad.This thesis mainly focused on the research of lung CT and CXR images,aiming to develop accurate segmentation methods for helping radiologists in diagnosis and treatment.Research contents are shown below:(1)Considering that the limited data would decrease the accuracy of the deep learning based segmentation method and the manually delineated contours required excessive manpower,a semi-automatic lung segmentation algorithm in CT images consisting of a closed polygonal line algorithm and backpropagation neural network algorithm was proposed.Firstly,the algorithm used a few manually delineated points as input followed by the closed polygonal line algorithm for obtaining data sequence.Secondly,machine learning was used for training data sequence.Finally,a smooth lung contour can be obtained,and the semi-automatic lung segmentation in CT images was accomplished.In the experimental results,we used private and public datasets to test the performance of the proposed method qualitatively and quantitatively.(2)Considering that the overfitting may happen during the backpropagation neural network's training,a semi-automatic lung segmentation method in CXR images based on hull and closed polygonal line algorithm was proposed.The hull algorithm consisting of the convex hull and concave hull algorithm was used to complete coarse segmentation,and then data augmentation can be achieved to avoid the overfitting that happened during the backpropagation neural network's training.Secondly,the optimization step was used for refining the coarse segmentation results.In the experimental part,we used more evaluation parameters to comprehensively prove the performance of the proposed method.Meanwhile,the local magnifications of both the apical region and costophrenic angle region were shown.(3)Considering that the high costs and difficulties in a manual intervention during the process of the semi-automatic segmentation,a fully automatic segmentation method called Mask Region-based Convolutional neural network(Mask-RCNN)algorithm and closed principal curve algorithm was developed.The performance of the proposed method was validated in several public lung datasets and compared with several existing methods.The experimental results showed that the proposed method had accurate and efficient performance for lung segmentation in CXR images.(4)Considering that the traditional closed principal curve based method was often affected by the abnormal vertices and the traditional machine learning without optimal model was not able to obtain the optimal global results during training,an automatic method combining with the enhanced machine learning and improved closed polygonal line method was proposed.The method was validated on several public datasets containing lesions and showed good clinical significance.Furthermore,the performance of the proposed method was proved via the quantitative and qualitative experiments.