Investigating computer aided diagnosis: Medical image analysis in CT colonography.

摘要

Computer-aided diagnosis (CAD) is generally defined as the diagnosis made by a radiologist or a physician who utilizes computer output as second opinion. This definition frames a practical work flow of how the computer should be used in medical diagnosis. The computer algorithms, as bases of CAD schemes, currently rely on the techniques of image processing, computer vision, pattern recognition, artificial intelligence, and other computer decision processes. An overview of these techniques and their applications in the medical environment, especially for CAD are presented in this thesis. Yet, CAD is far beyond a purely technical effort, and a variety of clinical and other issues must be addressed before it could be widely accepted in routine practices. Most CAD schemes are task specific; in recent years, breast cancer, lung cancer, and colon cancer detection are the three major early screening applications that attract most CAD investigators. After an introduction to CAD research, this thesis focuses on the technical issues; including the system design, implementation, and evaluation of CAD applied to CT colonography (CTC).; Colorectal cancer is the second leading cause for cancer related death in the united States, yet is the most preventable cancer if effective early screening is achieved. CTC is a promising alternative to the expensive and painstaking procedure of optical colonoscopy, which is the gold standard for colorectal cancer screening. A computerized polyp detection system is designed and implemented, which inputs CTC images; performs 3-D segmentation, geometric feature extraction, and pattern classification; and outputs a list of detected polyps. In colon lumen segmentation, a knowledge-based 3-D approach is developed and proven to be efficient and effective by a carefully designed evaluation scheme. In subsequent polyp detection, a rule-based classifier and a maximum a posteriori (MAP) classifier are developed for the situation of insufficient number of samples with possibly incorrect labeling under the special arena of CTC CAD. The detection results of the patient-level sensitivities and specificities are encouraging. Performance comparison among different classifiers is given by receiver operating characteristic (ROC) analysis and free response ROC (FROC) analysis.; In modern CAD research, a large number of diagnostic images are involved, numerous features are quantified, and various classifiers are employed. We have set up a research platform that efficiently manages this complexity. The platform features low-cost implementation, high-performance computation, database-powered management, and web-enabled interactive reporting. A convenient result presentation and an effective communication between engineers and physicians are established. We foresee an efficient platform for human-computer interaction and anticipate the same idea expanded to other modern CAD applications.