Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Amini, Amir A.

Committee Co-Chair (if applicable)

Frigui, Hichem

Committee Member

Park, Juw Won

Committee Member

Altiparmak, Nihat


Lung cancer is a leading cause of cancer death in the world. Key to survival of patients is early diagnosis. Studies have demonstrated that screening high risk patients with Low-dose Computed Tomography (CT) is invaluable for reducing morbidity and mortality. Computer Aided Diagnosis (CADx) systems can assist radiologists and care providers in reading and analyzing lung CT images to segment, classify, and keep track of nodules for signs of cancer. In this thesis, we propose a CADx system for this purpose. To predict lung nodule malignancy, we propose a new deep learning framework that combines Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) to learn best in-plane and inter-slice visual features for diagnostic nodule classification. Since a nodule's volumetric growth and shape variation over a period of time may reveal information regarding the malignancy of nodule, separately, a dictionary learning based approach is proposed to segment the nodule's shape at two time points from two scans, one year apart. The output of a CNN classifier trained to learn visual appearance of malignant nodules is then combined with the derived measures of shape change and volumetric growth in assigning a probability of malignancy to the nodule. Due to the limited number of available CT scans of benign and malignant nodules in the image database from the National Lung Screening Trial (NLST), we chose to initially train a deep neural network on the larger LUNA16 Challenge database which was built for the purpose of eliminating false positives from detected nodules in thoracic CT scans. Discriminative features that were learned in this application were transferred to predict malignancy. The algorithm for segmenting nodule shapes in serial CT scans utilizes a sparse combination of training shapes (SCoTS). This algorithm captures a sparse representation of a shape in input data through a linear span of previously delineated shapes in a training repository. The model updates shape prior over level set iterations and captures variabilities in shapes by a sparse combination of the training data. The level set evolution is therefore driven by a data term as well as a term capturing valid prior shapes. During evolution, the shape prior influence is adjusted based on shape reconstruction, with the assigned weight determined from the degree of sparsity of the representation. The discriminative nature of sparse representation, affords us the opportunity to compare nodules' variations in consecutive time points and to predict malignancy. Experimental validations of the proposed segmentation algorithm have been demonstrated on 542 3-D lung nodule data from the LIDC-IDRI database which includes radiologist delineated nodule boundaries. The effectiveness of the proposed deep learning and dictionary learning architectures for malignancy prediction have been demonstrated on CT data from 370 biopsied subjects collected from the NLST database. Each subject in this database had at least two serial CT scans at two separate time points one year apart. The proposed RNN CAD system achieved an ROC Area Under the Curve (AUC) of 0.87, when validated on CT data from nodules at second sequential time point and 0.83 based on dictionary learning method; however, when nodule shape change and appearance were combined, the classifier performance improved to AUC=0.89.