Classification of Small Bowel Lesions in Video Capsule Endoscopy Images Using Enhanced Deep Learning Model

Video capsule endoscopy (VCE) is a noninvasive procedure for diagnosing small bowel (SB) lesions. During VCE procedures, the miniature camera captures thousands of images, requiring considerable time and effort from healthcare professionals to examine each image for abnormalities. This study aims to develop and evaluate a modified DenseNet-201 convolutional neural network (CNN) model to assist in the automatic classification of SB lesions in VCE images. A representative dataset of 5,899 images was created by merging two state-of-the-art datasets, including six types of lesions and one class without lesions. A synthetic oversampling approach generated 1,273 synthetic images for the underrepresented class. Synthetic image quality was evaluated using texture metrics. The experiment was conducted on a set of 7,172 images using data partitions of 70 %, 15 %, and 15 % for training, validation, and testing, respectively, with a Monte Carlo cross-validation to verify the consistency of the experimental results. The model achieved an accuracy of 89.78 %, a precision of 89.66 %, a sensitivity of 89.69 %, a specificity of 89.72 %, and an F1 score of 89.67 % in the test set. The modified DenseNet-201 CNN model could be a valuable tool for diagnosing SB conditions and improving patient outcomes.