3-D IMAGE ANALYSIS TECHNIQUES FOR LUNG CANCER DETECTION ON COMPUTED TOMOGRAPHY (CT) SCANS

Abstract

Compelling clinical studies have shown that lung cancer screening allows for early diagnosis and treatment of lung cancer. Despite the high quality CT images available to the radiologists for decision making, missing lung cancer on the chest radiograph attributes to errors made by the radiologists in determining whether a lung nodule is malignant or not, on the basis of medical imaging and other clinical information. Several research studies have recommended Computer Aided Diagnosis (CADx) systems which facilitate early detection of lung cancer nodules at more curable stages. The significant and on-going research areas in CADx systems include automated segmentation systems and classification models for nodules characterization. The non availability of large databases has been a limitation for the training and development of accurate diagnostic models. The thesis aims to develop image analysis based methods in 3-D to improve the performance of CAD systems for lung cancer detection and pathological stage estimation. The contributed research work was carried out in the areas ol segmentation and cancer diagnostic models respectively which are integral parts of a CAD system. During the first phase of research, a semi-automatic 3-D segmentation algorithm developed for varying nodules sizes and shapes which enables accurate volumetric assessment and further analysis for nodules characterization in Computed Tomography (CT) images. The proposed model performs nodule segmentation by incorporating the geometric and region based properties of a lung nodule in CT images. An adaptive approach was developed to compute the mean gray level intensity of a lung nodule which is used as a stopping function besides the gradient properties to enable accurate nodule extraction from the CT images. The algorithm was tested and validated on multiple databases including LIDC, FDA, RIDER and CUMC with a mean spatial was overlap of 85% As second research contribution, a Bayesian Inversion framework was devised to predict the cancer occurrence as well as its pathological stage. The proposed stochastic model was tested on NLST database, a repository of low dose longitudinal CT data sets and is aimed to address the constraint of limited number of annotated CT datasets in lung cancer for multi-class classification. To develop the proposed framework, quantitative imaging features known as radiomic features were extracted from the segmented nodules (benign and malignant), analyzed and ranked using an intuitive approach to identify the diagnostic features for cancer and its stage estimation respectively. The carried out research work showed that Surface Area and Sum Entropy are the two radiomic features that possess general phenotype/characteristics to differentiate between benign and cancerous tumors of lung, colon and head and neck cancer respectively. These highly discriminating features were validated by integration into non-linear mathematical equation and achieved an accuracy of 97.0% for lung nodule classification. Another group of radiomic features was investigated in malignant nodules which can differentiate between early stage (stage I and stage II) and advanced stage (stage III and stage IV) cancer. Based on the analysis performed using regularization method and supervised feature selection, radiomic features which showed strong discrimination between early and advanced stage cancer were Sphericity, Large Dependence High Gray Level Emphasis (LDHGLE), Cluster Prominence, Small Area Emphasis and Strength respectively. These features were integrated into a mathematical equation for validation and achieved an accuracy of 86.84% when tested on data sets of malignant nodules with early and advanced stage cancer. By incorporating the aforementioned formulated equations for cancer and its stage detection respectively, the proposed stochastic framework predicted the occurrence as well as its pathological stage with an accuracy of 86%