Max-AUC feature selection in computer-aided detection of polyps in CT colonography

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Computerized detection of polyps in CT colonography (CTC) is crucial for colorectal cancer screening.
• Existing feature selection methods, like Wilks' lambda, may not optimize classification performance for polyp detection.
• Support vector machine (SVM) classifiers are widely used in medical image analysis.

Purpose of the Study:

• To develop and evaluate a novel feature selection method to enhance polyp detection performance in CTC.
• To compare the proposed method against conventional feature selection techniques.
• To improve the accuracy of computer-aided detection (CADe) schemes for colonic polyps.

Main Methods:

• A sequential forward floating selection (SFFS) procedure was employed for feature selection.
• The SFFS method aimed to maximize the area under the receiver operating characteristic curve (AUC) for classification performance.
• Two variants of the SFFS method were proposed, differing in their stopping criteria.
• The proposed methods were evaluated using a colonic-polyp database and compared against the Wilks' lambda stepwise method.

Main Results:

• The proposed SFFS feature selection method, coupled with a nonlinear SVM classifier, achieved 96% by-polyp sensitivity.
• The two variants of the SFFS method selected 29 and 7 features, respectively.
• The proposed method significantly reduced false-positive rates (4.1 and 6.5 per patient) compared to Wilks' lambda (18.0 per patient) at the same sensitivity level.

Conclusions:

• The proposed SFFS-based feature selection method significantly improves the performance of classifiers for polyp detection in CTC.
• This approach offers a more effective way to select relevant features for computer-aided detection (CADe) systems.
• The method demonstrates potential for enhancing the accuracy and efficiency of colorectal cancer screening through improved polyp identification.