[Paper Review] Feature Based Framework to Detect Diseases, Tumor, and Bleeding in Wireless Capsule Endoscopy

Omid Haji Maghsoudi, Mahdi Alizadeh|arXiv (Cornell University)|Jan 27, 2018

Gastrointestinal Bleeding Diagnosis and Treatment3 references3 citations

TL;DR

This paper proposes a feature-based framework for automated detection of tumors, bleeding, and other gastrointestinal abnormalities in wireless capsule endoscopy (WCE) frames using a combination of color, texture, and geometric features. The method employs SVM and neural network classifiers on segmented regions, achieving 100% sensitivity for tumor detection in frame-based analysis and 97% sensitivity for bleeding in pixel-based segmentation, significantly improving upon prior methods by integrating multi-feature analysis for broader lesion detection.

ABSTRACT

Studying animal locomotion improves our understanding of motor control and aids in the treatment of motor impairment. Mice are a premier model of human disease and are the model system of choice for much of basic neuroscience. High frame rates (250 Hz) are needed to quantify the kinematics of these running rodents. Manual tracking, especially for multiple markers, becomes time-consuming and impossible. Therefore, an automated method is necessary. We propose a method to track the paws of the animal in the following manner: first, segmenting all the possible paws based on color; second, classifying the segmented objects using a support vector machine (SVM) and neural network (NN); third, classifying the objects using the kinematic features of the running animal, coupled with texture features from earlier frames; and finally, detecting and handling collisions to assure the correctness of labelled paws. The proposed method is validated in sixty 1,000 frame video sequences (4 seconds) captured by four cameras from five mice. The total sensitivity for tracking of the front and hind paw is 99.70% using the SVM classifier and 99.76% using the NN classifier. In addition, we show the feasibility of 3D reconstruction using the four camera system.

Motivation & Objective

• Address the lack of comprehensive automated detection systems for multiple GI abnormalities beyond bleeding and ulcers in WCE.
• Overcome limitations of prior methods that relied on single-feature types (e.g., color only) and were specific to one lesion type.
• Develop a unified framework capable of detecting three major abnormal classes: tumors, bleeding, and five other diseases (e.g., Crohn’s, lymphangiectasia).
• Improve detection accuracy and robustness by integrating diverse features—color, texture (LBP, GLCM, Gabor), and geometric features—across frame-based and pixel-based analysis.
• Enable clinical utility by reducing physician review time through automated, accurate lesion localization and classification in WCE videos.

Proposed method

• Segment candidate abnormal regions in WCE frames using intensity and color-based thresholding to isolate potential lesions.
• Extract multi-scale features: color (RGB, HSI, LUV), texture (LBP, GLCM, Gabor filters, Law’s features), and geometric features (shape, size, orientation).
• Apply feature selection using correlation and Fisher score to reduce redundancy and select optimal features for each abnormality class.
• Train and evaluate two classifiers—Support Vector Machine (SVM) and Neural Network (NN)—on selected features for frame-based classification of tumor, bleeding, and other diseases.
• Implement a pixel-based segmentation method using selected texture and color features to localize abnormal regions within frames, with post-processing to refine boundaries.
• Integrate collision detection logic to resolve ambiguities in tracking and labeling, especially in overlapping or occluded regions.

Experimental results

Research questions

• RQ1 Can a multi-feature fusion approach combining color, texture, and geometric features significantly improve detection accuracy for multiple GI abnormalities in WCE compared to single-feature methods?
• RQ2 How does the performance of SVM and neural network classifiers compare when applied to frame-based detection of tumors, bleeding, and other lesions in WCE?
• RQ3 To what extent can a pixel-based segmentation method achieve high sensitivity and specificity in localizing abnormal regions, particularly for bleeding and tumors?
• RQ4 Can the proposed framework generalize across diverse lesion types (e.g., Crohn’s disease, lymphoid hyperplasia, xanthoma) beyond bleeding and ulcers?
• RQ5 How does the integration of temporal kinematic features and texture from previous frames enhance the robustness of abnormal region detection?

Key findings

• The frame-based method achieved 100% sensitivity for tumor detection, significantly outperforming previous methods that reported sensitivities between 85% and 88%.
• The pixel-based method for bleeding detection achieved 97% sensitivity, 97% specificity, and 94% precision—surpassing prior studies that reported lower specificity and precision.
• For the detection of other abnormalities (e.g., Crohn’s, lymphangiectasia), the pixel-based method achieved 96% sensitivity, 93% specificity, and 85% precision, demonstrating strong performance across diverse lesion types.
• Feature selection reduced the number of input features from 110 to 3 for bleeding, 12 for tumors, and 10 for other diseases, improving computational efficiency without sacrificing accuracy.
• The use of combined color, texture, and geometric features led to higher detection rates than methods relying on single-feature types, confirming the importance of multi-modal feature integration.
• The framework outperformed existing methods in both frame-based and pixel-based lesion classification, as validated by comparison tables showing superior sensitivity and specificity across multiple studies.