Random forest based computer-aided detection of polyps in CT colonography

As a promising second reader for computed tomographic colonography (CTC) screening, computer-aided detection (CAD) of colorectal polyps have been explored extensively. In this paper, we present a random forest (RF) based CAD scheme. First, a thick colon wall called volumetric mucosa was extracted by segmentation method from CTC images. We then computed the first and second order derivatives to perform the geometric analysis. Furthermore, the initial polyp candidates (IPCs) were detected by thresholding the geometric measurement. A set of features extracted from these IPCs were then fed into the RF classifier for both classification and false positive (FP) reduction. Finally, the detection results were presented as second opinions to the radiologists. The proposed RF-based CAD scheme was evaluated using two different datasets of patient studies, where the first dataset includes 49 CTC scans of 25 patients and the second dataset encompasses 86 scans of 53 patients. Using the RF based classification with feature selection, our presented CAD scheme achieved good performance, in particular, by leveraging the projection features computed in our previous method.