Bleeding detection from capsule endoscopy videos.

Area of Science:

• Medical imaging analysis
• Computer vision in healthcare

Background:

• Classifying disease signs in medical videos is challenging due to imbalanced datasets, leading to biased learning and low sensitivity.
• Conventional image classification methods struggle with the high class imbalance often found in medical video data.

Purpose of the Study:

• To develop and evaluate a robust classification method for identifying video frames containing bleeding lesions.
• To address the issue of data imbalance in medical video analysis for improved disease detection.

Main Methods:

• Implemented a data re-balancing technique involving over-sampling the minority class (bleeding lesions) and under-sampling the majority class.
• Constructed a Support Vector Machine (SVM) ensemble classifier utilizing re-balanced data and three distinct image features.
• Experimented with five datasets, each comprising approximately 55,000 images with a minority-to-majority class ratio of 1:145.

Main Results:

• The proposed method demonstrated superior sensitivity in detecting bleeding lesions compared to conventional approaches.
• Achieved comparative subjectivity with a slight improvement, indicating enhanced reliability of the classification.
• Preliminary results highlight the effectiveness of the re-balancing strategy and SVM ensemble for this specific task.

Conclusions:

• The developed classification method effectively handles imbalanced data for improved bleeding lesion detection in medical videos.
• This approach offers a promising solution for enhancing the sensitivity and accuracy of disease sign identification in clinical settings.
• Further research can explore broader applications of this technique in medical video analysis.