Novel 3D Radiomic Features for Computer-Aided Polyp Detection in CT Colonography

Abstract

State-of-the-art computer-aided detection (CAD) systems for colonic polyps in computed tomographic colonography (CTC) tend to yield high detection sensitivities with high false positive (FP) rates. This paper proposes a novel CTC CAD system using 3-D radiomic features that can obtain high detection sensitivity at low FP rate. Our previous shape-based methods are used to generate polyp candidates. Then, a wide variety of radiomic features are developed in the 3-D domain for each initial candidate, including the multiscale Weber local descriptors and the statistical descriptors calculated from various volumetric feature maps on the basis of computed tomography (CT) density, CT gradient, fractal dimension, curvature, and fast radial symmetry transform. These new 3-D radiomic features can characterize polyp candidates in terms of the shape, texture heterogeneity, and salient pattern. FP reduction is finally performed by Random Forests with a developed score rank method to tune the training set. 510 fluid-tagging CT scans from 255 patients with 130 polyps ≥5 mm were utilized to validate the proposed system in the fivefold cross-validation strategy. The detection result reached 98.5% by-polyp sensitivity at 2.0 FPs per scan for polyps ≥5 mm. Experimental results indicate that the proposed system yielded a detection performance with high sensitivity and low FP rate. We believe that the proposed system would assist radiologists in increasing the detection accuracy and reducing the interpretation time during colon cancer screening.

Keywords

• Computer-aided detection
• colonic polyp
• radiomic feature
• Random Forests