Detection of colonic polyps in computed tomographic images

Cancer is a leading cause of death and colorectal cancer is amongst the most common cancers worldwide. Colorectal cancer is also the third leading cause of cancer deaths in the United States. Colonic polyps are usually known to be a precursor to colorectal cancer. In western countries, over 95% of colorectal cancers arise from colonic polyps. Thus, early detection and treatment provide a higher cure rate, and also, the timely removal of precancerous polyps can prevent up to 90% of deaths by. Computed tomography colonography (CTC) is an inspection method for examining the inside of the colon structure. This method enables radiologists to observe the inner side of the colons using abdominal CT images. CTC method significantly reduces the burden of colonoscopy procedures. CTC is also less painful than other inspection methods. CTC method allows a rapid structural assessment of the whole colon without application of sedatives and minimizes the risks of procedure-related complications. Nowadays, Computer Aided Detection (CAD) systems are used in order to help radiologists to detect colonic polyps over CT scans. It is possible to reduce the detection time and increase the detection accuracy rates by using CAD systems. In this study, a novel CAD system for automated detection of colonic polyps in CT scans was developed. To evaluate the performance of the CAD system, supine CT scans were performed on 30 patients. The proposed CAD system is a multistage implementation whose main components are: automatic colon segmentation, candidate surface extraction, feature extraction and classification. Segmentation procedure can be defined briefly as the extraction of colonic structure through CT scans. In this study, depends on the presence of contrast material a segmentation procedure which can handle both air and fluid filled parts of the colon separately were developed. To minimize the noise artifacts that may be caused by segmentation, the colon structure was re-segmented with fuzzy C-means (FCM) algorithm to generate its latest form. When the segmentation algorithm outputs were inspected, the reconstruction of the colon in the collapsed CT data (the colon collapse into a series of disconnected segments due to the presence of blockages caused by the residual fecal material) was accomplished. The proposed segmentation algorithm was successfully determined the colon structure for all patients and it was observed that all colonic polyps were preserved within the segmented structures. Colonic polyps can be described as an abnormal growth which arises on the inner surface of the colon and in the CT data; polyps may appear roughly as an elliptical and blob-like structures. This distinctive shape features may use to differentiate polyps, folds and normal colonic wall. In the literature, many different approaches use the shape features to detect polyp candidates. In this thesis, Marr-Hildreth (LoG) algorithm was employed to extract the elliptical structures as potential polyp candidate regions. The proposed polyp detection method successfully determined all colonic polyps. Also, in order to reduce false positive rate, we fit primitive shapes such as circle, line and sphere and calculate the residual of each least square solution. The final component of the CAD system was generated a two dimensional projection image for each candidate and subsequently, nine morphological discriminative features were extracted from the projection images. These morphological discriminative features were then used for polyp classification using a committee of multi layer perceptron (MLP) classifiers. Our CAD system performs 94,59% of sensitivity at 3,9 false positives per dataset.