Performance assessment of a bleeding detection algorithm for endoscopic video based on classifier fusion method and exhaustive feature selection. (February 2018)

• Record Type: Journal Article
• Authors: Deeba, Farah; Islam, Monzurul; Bui, Francis M.; Wahid, Khan A.
• Is Part Of: Biomedical signal processing and control. Volume 40(2018)
• Page Start: 415
• Page End: 424
• Publication Date: 2018-02
• Subjects: Capsule endoscopy -- Color features -- Automated bleeding detection -- Classifier fusion -- Nested cross validation -- SVM score; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2017.10.011
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: An efficient bleeding detection system has been proposed based on a classifier fusion algorithm. Optimum training has been ensured by adopting a nested cross validation strategy. Robustness of the system has been improved by fusing multiple classifiers based on SVM score. Comparison with state-of-the-art algorithms validates the superiority of the proposed method for diverse dataset. Abstract: Capsule Endoscopy (CE) is a non-invasive clinical procedure that allows examination of the entire gastrointestinal tract including parts of small intestine beyond the scope of conventional endoscope. It requires computer-aided approach for the assessment of video frames to reduce diagnosis time. This paper presents a computer-assisted method based on a classifier fusion algorithm which combines two optimized Support Vector Machine (SVM) classifiers to automatically detect bleeding regions present in CE frames. The classifiers are based on RGB and HSV color spaces; the image regions are characterized on the basis of statistical features derived from the first-order histogram probability of respective color channels. A nested cross validation strategy has been adopted for the parameter tuning and feature selection to optimize the classifiers. The optimum feature sets for the best performance are evaluated after exhaustive analysis. The proposed fusion approach achieves an average accuracy of 95%, sensitivity of 94% and specificity of 95.3% for a dataset of 8872 CE frames, which is higher than that obtained from a single classifier. Comparison with the state-of-the-art algorithms exhibits that the proposed method yields superior performance for diverse dataset.