Research On Multi-stage Method For Pulmonary Embolism Detection From CT Images Based On Deep Learning

Pulmonary embolism(PE)refers to a sudden blockage in a pulmonary artery.It is a disease with high morbidity and mortality.Timely diagnosis and active treatment can effectively reduce the mortality of patients.Nowadays,CT Pulmonary Angiography(CTPA)is the main method to diagnose pulmonary embolism in clinical practice.In a CT image,pulmonary embolism appears as a dark area in the middle of a bright area.However,there are often many dark areas with similar characteristics in a CT image,which will bring a certain degree of difficulty for radiologists to interpret CTs.Manually interpreting CTs is sensitive to radiologists’ experience and eye fatigue,which is also a time-consuming process.Therefore,a computer aided detection system with good performance will effectively help the radiologists in their work.Existing methods for PE detection utilize handcrafted features to perform candidate proposal and false positive removal.However,due to the limited representation ability of these handcrafted features,these methods often suffer from a high false positive rate.To solve this problem,we are committed to designing a pulmonary embolism detection method based on deep learning,so that the algorithm can achieve higher sensitivity while generating a relatively low number of false positives.Following the motivation of candidate proposal and false positive removal,the pulmonary embolism detection method proposed in this thesis consists of three stages: candidate extraction based on 3D convolutional neural network(CNN),false positive(FP)removal based on 2D CNN,and FP removal based on artery/vein separation.The contribution of our method are as follows:(1)A combination of 3D CNN and 2D CNN,which full utilizes the ability of 3D CNN to extract 3D spatial information and the advantage of 2D CNN to reduce the number of parameters.(2)A novel image representation is realized for each candidate,which effectively reduces the 3D data to 2D while retaining the characteristics of the blood vessel.(3)A subtree separation algorithm is designed to extract lung vessel subtrees,and the prediction results of neural network are corrected in these subtrees to obtain consistent A/V separation results.Our method is evaluated in the PE Challenge dataset consisting of 20 testing cases and PE 129 dataset collected from local hospital consisting of 29 testing cases,respectively.On PE challenge dataset,our method achieves a sensitivity of 75.4% at 2 false positives per scan at 08)8)localization error,which is superior to the winning system in the literature.On PE129 dataset,our method achieves sensitivities of 76.3%,78.9% and 84.2% at 2 false positives per scan at 0mm,2mm and 5mm localization error,respectively.