Enhancing Gastrointestinal Diseases Classification Using Gastroenterology Imaging and Transfer Learning

Abstract

Gastrointestinal (GI) diseases present a significant global health challenge, requiring the advancement of enhanced diagnostic tools for accurate and efficient classification. This study explores the enhancement of GI disease classification through the application of transfer learning, utilizing various deep learning architectures, including ResNeSt50, ResNeSt101, ResNeSt200, ResNet50, and VGG16, applied to the Kvasir V2 dataset and a local prepared Ethiopian dataset. Initially, various models were evaluated using the Kvasir V2 dataset, with ResNeSt50 emerging as the top performer. The best results with ResNeSt50 were achieved using a configuration with a learning rate of 0.01, a dropout rate of 0.5, a weight decay of 1e-3, unfreezing of the last two layers and FCL, scheduler factor of 0.2. This setup achieved a validation set accuracy of 93.33%, a test set accuracy of 93.22%, and a local test set accuracy of 88.75%. To further enhance the model’s robustness and accuracy, the Kvasir V2 dataset was incorporated with our local dataset during training. With this combined dataset, ResNeSt50 achieved a validation accuracy of 93.22% and a test accuracy of 93.11%. The model also maintained consistently high metrics across F1 score, precision, and recall, all within the range of 93.32% to 93.39%. This incorporation of the local dataset significantly improved the local test set accuracy from 88.75% to 93.17%, highlighting its strong generalization capabilities across different datasets. The findings underscore the effectiveness of the ResNeSt50 model, particularly with these specific hyperparameter settings and data augmentation strategies, in enhancing GI disease classification. This work emphasizes the power of transfer learning in improving diagnostic accuracy and provides and establishes a basis for future research, including the expansion of datasets and the exploration of other deep learning architectures.