Automated Segmentation And Classification Of Pulmonary Embolism And Nodules Using Neural Networks

Pulmonary Embolism(PE)is the most frequent cause of cancer-related mortality in humans.Computed tomography(CT)for medical disease screening is an efficient form of reliable and early detection that greatly increases the survival rate.However,interpreting medical images and creating decisions for evaluation or care involve specifically qualified medical experts.Present method of interpreting diagnostic images is labor-intensive,time-consuming,expensive and error-prone.It will be more appropriate to provide a computer-aided model that would automatically render suggestions for diagnosis.Recent developments in deep learning encourage us to reconsider the way clinical diagnosis is focused on medical images.Early detection has been proven to be crucial to provide patients with the greatest likelihood of rehabilitation and survival.In this thesis,we present a completely automated segmentation and classification of pulmonary embolism and nodule using neural network framework based on CT images.Our proposed work consists of two parts,PE segmentation(preprocessing and train a model to do segmentation of PE)and classification(diagnose and classifying candidate nodules as benign or malignant).For PE segmentation,we designed a new effective image pre-processing method combined with CT window technique and image cropping.this proposed model established is an encoder-decoder convolutional network.The residual block instead of the original convolutional block were used for U-Nets.In order to do classification,of nodules,two deep 3D Conv Nets were built for nodule detection and classification,respectively.Furthermore,we validate a function mixing nodule scale with raw 3D cropped nodule pixels,and employ Gradient Boosting Machine as a classifier,as well as achieve outperform accuracy results as compare to traditional classifiers.We use 3D dual path networks(DPN)as components.Precisely,based on the efficiency of Faster R-CNN for object detection.we suggest 3D Faster R-CNN for nodule detection based on 3D dual path framework and U-net-like encoder-decoder layout,as well as deep 3D dual path network for nodule classification.Finally,we achieve outperform accuracy output in both proposed work,which demonstrates the accuracy of neural network features.