New classifier architecture and training methodologies for lung nodule detection in chest radiographs and computed tomography
Date of Award
Ph.D. in Electrical and Computer Engineering
Department of Electrical and Computer Engineering
Advisor: Russell C. Hardie
Early detection of pulmonary lung nodules plays a significant role in the diagnosis of lung cancer. Radiologists use Computed Tomography (CT) and Chest Radiographs (CRs) to detect such nodules. In this research, we propose various pattern recognition algorithms to enhance the classification performance of the Computer Aided Detection (CAD) system for lung nodule detection in both modalities. We propose a novel optimized method of feature selection for clustering that would aid the performance of the classifier. We make use of an independent CR database for training purposes. Testing is implemented on a publicly available database created by the Standard Digital Image Database Project Team of the Scientific Committee of the Japanese Society of Radiological Technology (JRST). The JRST database comprises 154 CRs containing one radiologist confirmed nodule in each. We make use of 107 CT scans from publicly available dataset created by Lung Image Database Consortium (LIDC) for this study. We compare the performance of the cluster-classifier architecture to a single aggregate classifier architecture. Overall, with a specificity of 3 false positives per case on an average, we show a classifier performance boost of 7.7% for CRs and 5.0% for CT scans when compared to single aggregate classifier architecture. Furthermore, we study the performance of a CAD system in CT scans as a function of slice thickness. We believe this study has implication for how CT is acquired, processed and stored. We make use of CT cases acquired at a thickness of 1.25mm from the publicly available Lung Nodule Analysis 2016 (LUNA16) dataset for this research. We study the CAD performance at a native thickness of 1.25mm and various other down-sampled stages. Our study indicates that CAD performance at 2.5mm is comparable to 1.25mm and is much better than at higher thicknesses. In addition, we propose and compare three different training methodologies for utilizing non-homogenous thickness training (i.e., composed of cases with different slice thicknesses). We utilize cases acquired at 1.25mm and 2.5mm respectively from the LUNA16 dataset for this study. These methods include: (1) aggregate training using the entire suite of data at their native thickness, (2) homogeneous subset training that uses the subset of training data that matches each testing case; and (3) resampling all training and testing cases to a common thickness. Our experimental results indicate that resampling all training and testing cases to 2.5mm provides the best performance among the three training methods compared. Furthermore, the resampled 2.5mm data require less memory and process faster than the 1.25mm data.
Diagnostic imaging Data processing, Pattern recognition systems Data processing, Lungs Biopsy Imaging, Lungs Diseases Imaging, Electrical Engineering, Medical Imaging, Biomedical Research, Computer aided detection, Computed tomography, Chest radiographs, Lung nodules, Slice Thickness
Copyright 2017, author
Narayanan, Barath Narayanan, "New classifier architecture and training methodologies for lung nodule detection in chest radiographs and computed tomography" (2017). Graduate Theses and Dissertations. 1335.