Computer-Aided Diagnosis Of Lung Nodules And Liver Nodules With Radiomic Approach In CT Images

Cancers seriously endanger human lives in worldwide.Among people who died in cancers,the number of lung cancer deaths was the highest,and the number of liver cancer deaths ranked fourth(after lung cancer,colorectal cancer,and stomach cancer).Early diagnosis of cancer has important implications for patients.Both lung cancer and liver cancer are solid cancers and they all demonstrate nodular lesions in radiology findings.Early diagnosis with CT images both can improve the prognosis of patients and has a great value in clinical practice,however,it is also very challenging at the same time.The computer-aided lung nodule detection contains many false positive detections which are mostly from the pulmonary vascular bifurcations.In this dissertation,the selective enhancement features were applied to reduce the false positive detections.The selective enhancement features can quantitively analyze spherical enhancement images for lung nodules and tubular enhancement images for blood vessels to reduce false positive detections.The selective enhancement features combined with the radiomic features and the random forest classifier can improve the sensitivity and reduce false positive for lung nodule detections and can help radiologists screen for lung cancer.After lung nodules are found in the images,efficient and accurate diagnosis of lung nodules to determine whether they are benign or malignant is critical.In this dissertation,the selective enhancement features were utilized to extract the vessel convergence sign which is an imaging phenotype of lung cancer.The radiomic features were applied to extract heterogeneity information and multi-frequency characteristics of lung nodules.All the features were analyzed by a random forest classifier to distinguish the benign and malignant lung nodules and achieved a good performance.This method is valuable for diagnosis of lung nodules with a single CT scan.It is a noninvasive and low-cost method for clinical decision support in lung cancer diagnosis and screening.This dissertation continued to explore the application of computer-aided diagnosis in liver nodules.The difference between liver nodules and lung nodules brings new challenges to our research.In clinical practice,radiologists can diagnose liver nodules based on comparative analysis between arterial phase CT images and portal venous phase CT images,but this method is ineffective for indeterminate liver nodules without quasi-pathognomonic imaging phenotypes.In addition,the non-optimal acquisitions of enhanced CT images will sometimes make it more difficult to diagnose indeterminate liver nodules.In order to solve this problem,this dissertation proposed an enhanced CT trigger timing factor combined with radiomic features to analyze the changes of the lesion between the portal and arterial phases,which can effectively diagnose indeterminate liver nodules.Furthermore,obtaining trigger timing factors was time-consuming and labor-intensive,therefore,a fully automatic method based on dual input convolutional neural network with the aorta and portal vein images was proposed to obtain trigger timing factors fast.This method eliminated the interaction of radiologists and significantly shortened the analysis time.This dissertation studied computer-aided diagnosis techniques for lung nodules and liver nodules and proposed selective enhancement features and trigger timing factors with radiomics to analyze chest CT images and abdominal multi-phase enhanced CT images.These techniques can help radiologists to diagnose lung nodules and indeterminate liver nodules effectively and provide incremental values for the early diagnosis of lung cancers and liver cancers and benefit the patients.