Geometry, Imaging and Computing

Volume 2 (2015)

Number 1

Adaptive kernel based multiple kernel learning for computer-aided polyp detection in CT colonography

Pages: 23 – 45

DOI: https://dx.doi.org/10.4310/GIC.2015.v2.n1.a2

Authors

Ming Ma (Department of Computer Science, Stony Brook University, Stony Brook, New York, U.S.A.)

Lihong Li (Department of Engineering Science and Physics, City University of New York, Staten Island, N.Y., U.S.A.)

Hao Han (Department of Radiology, Stony Brook University, Stony Brook, New York, U.S.A.)

Yifan Hu (Department of Radiology, Stony Brook University, Stony Brook, New York, U.S.A.)

Xianfeng Gu (Department of Computer Science, Stony Brook University, Stony Brook, New York, U.S.A.)

Zhengrong Liang (Department of Radiology, Stony Brook University, Stony Brook, New York, U.S.A.)

Abstract

Computer-aided detection (CAD) of colonic polyps, as a second reader for computed tomographic colonography (CTC) screening, has earned extensive research interest over the past decades. False positive (FP) reduction in the CAD system plays a crucial role in detecting the polyps. To improve the performance of FP reduction and better assist the physician’s diagnosis, we propose an adaptive kernel based multiple kernel learning (MKL) method for CAD of colonic polyps, called AK-MKL. This method builds a more adaptive synthesized classifier by incorporating an adaptive kernel into a set of predefined base kernels for better performance in differentiating true polyps from FPs, which is implemented by learning an optimal combination of a collection of those kernel-based classifiers. Performance evaluation for the presented AK-MKL method was performed on a CTC database, consisting of 25 patients with 50 CT scans. In terms of the AUC (area under the curve of receiver operating characteristic) and accuracy merits, the experimental results showed that our AK-MKL method achieves better performance, compared with two other different methods, i.e., one classifier based on support vector machine (SVM) with only one adaptive kernel (AK-SVM) and the other one based on multiple kernel learning only (MKL).

Full Text (PDF format)

Published 21 September 2015