Advanced gastrointestinal image classification based on vision transformer, hypercolumn and biological optimization

Medical image analysis plays a crucial role in the accurate diagnosis of gastrointestinal disorders. The Kvasir dataset, which consists of endoscopic images annotated by experienced medical professionals, serves as a valuable resource for developing advanced machine learning models. This study introduces a novel hybrid methodology that enhances classification performance by leveraging transformer-based models, deep feature extraction, and optimization techniques. Initially, three different Vision Transformer (ViT) models were employed for image classification, and their respective performance metrics were recorded. To further improve feature representation, the hypercolumn technique was integrated and trained, leading to additional classification results. Next, deep features were extracted from both standalone ViT models and hypercolumn-integrated models. These feature sets were then classified using Decision Trees, Random Forest, and Support Vector Machines. The two best-performing models were selected, and their feature representations were combined to construct a more comprehensive feature set. Feature selection was then performed using the Satin Bowerbird Optimization algorithm, refining the feature set to improve both efficiency and accuracy. With the proposed approach, a 96.88% overall accuracy was achieved. The proposed framework demonstrated strong internal classification performance on the Kvasir dataset and provides a methodological basis for transformer-based gastrointestinal image analysis. However, its generalizability and clinical utility remain to be established through independent, multicentre, and prospective validation studies.