Shape influence in medical image segmentation with application in computer aided diagnosis in CT colonography.

摘要

Computer-aided diagnosis (CAD) is a procedure in medicine that assists radiologists or physicians in the interpretation of medical images. The application of CAD in screening colorectal cancer (CRC) has been studied for more than two decades. CRC is the second most deadly form of cancer in men and women in the United States. Nearly all CRC arises from polyps and is preventable if polyps are removed in their early stage. Recent researches show that radiologists demonstrated higher accuracy in finding polyps with CAD than without CAD. In this dissertation, we review the current status of CAD research in CT Colonography (CTC) and propose a new method to segment and detect polyps in CTC CAD.;Advanced polyp segmentation and detection methods can improve the cost effectiveness of CTC by reducing the time used by radiologists in examining CTC studies. With polyp segmentation, we can characterize a polyp by its size, height, volume, and texture, all of which can be computed automatically. And this information is valuable in discriminating polyps from false positive detections such as haustral folds, residual stool, and the rectum tube.;We propose a model-based approach to segment and detect polyps. Initially, a number of manually segmented polyps are aligned to remove the translation, rotation, and scaling effects. A polyp shape model is constructed using the aligned polyps, among which the shape variances are captured by principal component analysis. A model-based registration method is employed to transform and deform the polyp shape model in order to match a polyp in a CTC study. In the final step, a polyp surface is identified by a deformable registration. We evaluated our segmentation results by comparing them to manually segmented polyp surfaces. The dice coefficient, which indicates the overlap ratio, is 84.5%+/-3.7 for 19 polyps in 12 patients. In order to improve the polyp detection in CAD, we derived a new feature, the magnitude of deformation from a polyp shape model to a polyp candidate, from the segmentation results. Since we have both a polyp model and a segmented polyp candidate, more features can be extracted to reduce false positive detections in the future.;We contributed to the CTC CAD field as follows: (1) devised a new approach to segment and detect polyps based on polyp shape models; (2) promoted the open source development in CTC CAD by utilizing 3D Slicer, ITK, and VTK; (3) improved the classification system design in CTC CAD by comparing a group of classifiers and tuning their parameters.