[Paper Review] Superpixels Based Segmentation and SVM Based Classification Method to Distinguish Five Diseases from Normal Regions in Wireless Capsule Endoscopy

Omid Haji Maghsoudi|arXiv (Cornell University)|Nov 17, 2017

Gastrointestinal Bleeding Diagnosis and Treatment26 references3 citations

TL;DR

This paper proposes a superpixel-based segmentation and SVM classification method to detect five gastrointestinal diseases—bleeding, Crohn’s disease, lymphangiectasia, xanthoma, and lymphoid hyperplasia—in wireless capsule endoscopy (WCE) images. Using SLIC and quick shift (QS) superpixels, the method extracts color and texture features for SVM classification, achieving 92–94% accuracy, 91–95% specificity, and 88–93% sensitivity, with SLIC being significantly faster than QS.

ABSTRACT

Wireless Capsule Endoscopy (WCE) is relatively a new technology to examine the entire GI trace. During an examination, it captures more than 55,000 frames. Reviewing all these images is time-consuming and prone to human error. It has been a challenge to develop intelligent methods assisting physicians to review the frames. The WCE frames are captured in 8-bit color depths which provides enough a color range to detect abnormalities. Here, superpixel based methods are proposed to segment five diseases including: bleeding, Crohn's disease, Lymphangiectasia, Xanthoma, and Lymphoid hyperplasia. Two superpixels methods are compared to provide semantic segmentation of these prolific diseases: simple linear iterative clustering (SLIC) and quick shift (QS). The segmented superpixels were classified into two classes (normal and abnormal) by support vector machine (SVM) using texture and color features. For both superpixel methods, the accuracy, specificity, sensitivity, and precision (SLIC, QS) were around 92%, 93%, 93%, and 88%, respectively. However, SLIC was dramatically faster than QS.

Motivation & Objective

• To develop an automated method for detecting five specific gastrointestinal diseases in WCE images.
• To address the time-consuming and error-prone manual review of over 55,000 WCE frames per study.
• To improve disease detection accuracy by leveraging superpixel-based segmentation and SVM classification.
• To compare the performance and efficiency of SLIC and quick shift superpixel algorithms in clinical image segmentation.

Proposed method

• Uses superpixel segmentation via SLIC and quick shift (QS) to group spatially coherent image regions.
• Extracts color and texture features from segmented superpixels for classification.
• Applies support vector machines (SVM) to classify regions as normal or abnormal based on extracted features.
• Employs HSI and RGB color spaces, along with Haralick, Gabor, and Law’s texture features for feature extraction.
• Evaluates performance using accuracy, specificity, sensitivity, and precision metrics across multiple superpixel counts.
• Compares SLIC and QS in terms of segmentation speed, accuracy, and robustness across disease types.

Experimental results

Research questions

• RQ1 How effective are SLIC and QS superpixel methods in segmenting five specific gastrointestinal diseases in WCE images?
• RQ2 What is the impact of varying superpixel count on classification accuracy and sensitivity for different diseases?
• RQ3 How does the proposed method compare to existing frame-based and pixel-based approaches in terms of detection performance?
• RQ4 Which superpixel algorithm—SLIC or QS—offers a better trade-off between accuracy and computational efficiency?
• RQ5 Can superpixel-based segmentation improve the detection of diseases with varying spatial extents, such as bleeding (large regions) versus xanthoma (small regions)?

Key findings

• SLIC and QS superpixel methods achieved 92–94% accuracy, 91–95% specificity, and 88–93% sensitivity in classifying five WCE diseases.
• SLIC demonstrated significantly faster processing time (0.7 seconds per frame) compared to QS, which was slower and more parameter-sensitive.
• For diseases with larger regions (e.g., bleeding, Crohn’s, lymphoid hyperplasia), performance declined beyond 100 superpixels due to over-segmentation.
• For smaller lesions (e.g., lymphangiectasia, xanthoma), performance improved with higher superpixel counts, indicating better resolution of small abnormalities.
• SLIC-based classification achieved 97% sensitivity and 97% specificity for bleeding, outperforming prior frame-based methods.
• The method showed superior performance compared to existing approaches, particularly in sensitivity and precision for lymphoid hyperplasia (93% sensitivity, 90% precision) and Crohn’s disease (92% sensitivity, 84% precision).