This paper presents an automated classification system for identifying abnormalities in Video Capsule Endoscopy (VCE) frames using the DINOv2 vision transformer model. Video capsule endoscopy is a vital non-invasive method for diagnosing gastrointestinal diseases, but manual frame-by-frame analysis can be time-consuming and subject to errors. To address this challenge, we developed a deep learning pipeline capable of classifying ten types of abnormalities, including angioectasia, bleeding, erosion, and polyps. Our model leverages the DINOv2 architecture for feature extraction, combined with fully connected layers for multi-class classification. The training dataset was augmented with random transformations to improve generalization, while test images were processed using standardized resizing and normalization techniques. The model was fine-tuned on over 37,000 labeled VCE frames and evaluated on a separate validation set. Results demonstrated a high overall accuracy of 91%, with particularly strong performance on the "Normal" class (F1-score of 0.97) and a balanced F1-score across other abnormality classes. Despite the dataset imbalance, the DINOv2 model achieved robust results with an average AUC of 0.85. This automated classification approach significantly reduces the need for timeintensive manual analysis and offers a scalable solution for assisting gastroenterologists in early detection of gastrointestinal diseases.
