Bleeding detection from capsule endoscopy videos

Reviewing medical videos for the presence of disease signs presents a unique problem to the conventional image classification tasks. The learning process based on imbalanced data set is heavily biased and tends to result in low sensitivity. In this article, we present a classification method for finding video frames that contain bleeding lesions. Our method re-balances the training samples by over-sampling the minority class and under-sampling the majority class. An SVM ensemble is then constructed using re-balanced data of three kinds of image features. Five sets of image frames were used in our experiments, each of which contains approximately 55,000 images and the ratio of minority and majority class is about 1:145. Our preliminary results demonstrated superior performance in sensitivity and comparative subjectivity with slight improvement.