Pulmonary Nodules Detection And Classification In Chest CT Using Deep Learning Techniques

Pulmonary cancer is one of the major cancers causing mortality throughout the world.However,the detection of lung cancer in its early stages can increase the patient’s survival rate and prove to be a vital lifesaving step.For lung cancer detection,numerous computer-assisted detection(CADe)and computer-assisted diagnosis(CADx)systems have been developed which use CT-Scan imaging modality.Various image processing techniques are part of conventional CADe systems which provide quantitative pulmonary nodule detection.The conventional method used by radiologists for detecting nodule presence in CT images is tedious and challenging;to assist in the ecient and accurate diagnosis process,various decision support systems have been developed over the years to provide a second opinion.Manual delineation of lung nodules by radiologists needs to be automatized by the computer-aided decision support system to make it more ecient and accurate.Recent advancement in deep learning techniques has enabled these CADx to automatically model high-level abstractions in CT-Scan images using multi-layered Convolutional Neural Network(CNN).Unlike the existing CAD systems that focused either on detection of nodules or classication of nodules,the proposed deep learning based CAD systems not only detects the nodules but also classies detected nodules as benign or malignant and identies the cancer stage.False Positive results in case of real world lung cancer is relatively very high,calculated to be approximately 42.86% according to the National Cancer Institute.The overall aim of the proposed CAD systems is to reduce the False Positive(FPs)per scan while improving the mean accuracy and sensitivity.The rst proposed model,DFCNet,is based on the Deep Fully Convolutional Neural Network which is used for classication of each detected pulmonary nodule into four lung cancer stages.Comparison of proposed classier is done with the existing CNN based CAD system.Overall accuracy of CNN,TumorNet and DFCNet was recorded as 83.64%,84.21% and 91.58%,respectively.A novel 3D Region-based Fully Convolutional Network(3D-RFCN)based automated decision support system for lung nodule detection and classication was proposed.This 3D-RFCN model is used as an image classier backbone for feature extraction along with the novel multi-Level Region Proposal Network(mLRPN)withposition-sensitive score maps(PSSM)being explored.Our proposed system has been trained and evaluated using LIDC dataset.Our 3D-RFCN model surpassed the state-of-the-art nodule detection and classication methods by achieving a sensitivity of 97.4% and classication accuracy of 97.61%.Our third proposed CAD system comprises of 3D residual U-Net for nodule detection.Initially,the 3D residual U-Net resulted in false positives results,therefore a novel False-Positive Reduction Algorithm(FPRA)was proposed for the improvement of nodule detection.The detected nodules are assigned a malignancy probability using Malignancy Score-Based Approach(MSBA)algorithm and classied nodules into four classes on the basis of its respective malignancy score.Our CAD system obtained a sensitivity of 97.65% and an average classication accuracy of 96.37% in extensive experimental results.Lastly,we developed a computer-aided decision support system for lung detection based on a 3D Deep Convolutional Neural Network(3DDCNN)architecture and introduced a novel multi-Region Proposal Network(mRPN)for automatic selection of potential region-of-interests.To further improve the eciency and performance of our proposed method,we integrated cloud computing into our CAD system.We trained and validated our Cloud-Based 3DDCNN on the datasets provided by Shanghai People’s Hospital No.6,as well as LUNA16,ANODE09,and LIDC-IDRI.Our proposed3 DDCNN model outperformed the existing methods with 98.7% sensitivity at 1.97 FPs per scan.The experimental evaluation of these proposed methods is performed on LUNA16,ANODE09,LIDC-IDRI,and multi-sample per subject clinical dataset from Shanghai People’s Hospital No.6.The performance evaluation results obtained from these CAD systems demonstrate the potential of deep learning,in combination with cloud computing,for accurate and ecient lung nodule detection using CT imaging,which could help doctors and radiologists in saving human lives.