Presentation Title: A Modular Approach to Lung Nodule Detection from Computed Tomography Images using Artificial Neural Networks and Content Based Image Representation

Committee:

• Dr.Jianhua Chen (Major Professor)
• Dr. Helmut Schneider (Minor Professor)
• Dr. Steven Bujenovic
• Dr. Bahadir Gunturk
• Dr. Suresh Rai(Dean's Representative)
• Dr. Kenneth L. (Kip) Matthews II
• Dr. Evangelos Triantaphyllou

Date: April 1, 2009
Time: 3:30 PM
Location: Highway Safety research Group

Abstract:
Lung cancer is one of the most lethal cancer types. Research in computer aided detection (CAD) and diagnosis (CADD) for lung cancer aims at providing effective tools to assist physicians in cancer diagnosis and treatment to save lives. In this dissertation, we focus on developing a CAD framework for automated lung cancer nodule detection from 3D lung CT images. Nodule detection is a challenging task that no machine intelligence can surpass human capability to date. On the other hand, human recognition power is limited by vision capacity and may suffer from work overload and fatigue, whereas automated nodule detection systems can complement human expert’s efforts to achieve better detection performance. The CAD framework proposed encompasses several desirable properties such as mimicking physicians by means of geometric multi-perspective analysis, computational efficiency, and the most importantly producing high performance in detection accuracy. As the central part of the framework, we develop a novel hierarchical modular decision engine implemented by Artificial Neural Networks (ANN). One advantage of this decision engine is that it supports the combination of spatial-level and feature-level information analysis in an efficient way. Our methodology overcomes some of the limitations of current lung nodule detection techniques by combining geometric multi-perspective analysis with global and local feature analysis. The proposed modular decision engine design is flexible to modifications in the decision modules; the engine structure can adopt the modifications without having to re-design the entire system. The engine can easily accommodate multi-learning scheme and parallel implementation so that each information type can be processed (in parallel) by the most adequate learning technique of its own. We have also developed a novel shape representation technique that is invariant under rigid-body transformation and we derived new features based on this shape representation for nodule detection. We implemented a prototype nodule detection system as a demonstration of the proposed framework. Experiments have been conducted to assess the performance of the proposed methodologies using real-world lung CT scan data. Several performance measures for detection accuracy are used in the assessment. The results show that the decision engine is able to classify patterns efficiently with very good classification performance.

All are invited.