Comparison of Convmixer Method and Resnet Method in Classification and Detection of Gastrointestinal Diseases Using Kvasir Dataset

Abstract

This research discusses the comparison of ConvMixer and ResNet methods in the classification and detection of gastrointestinal diseases using the Kvasir dataset. Gastrointestinal diseases are often difficult to detect early due to the similarity of visual patterns in endoscopy images, requiring an efficient deep learning-based solution. The purpose of this research is to compare and evaluate the models used. The research used a quantitative approach with an experimental method. Endoscopy image data was processed through augmentation, normalization, and division of the dataset into train, validation, and test. ConvMixer and ResNet were trained with customized hyperparameters, and evaluated using accuracy, precision, recall, and F1-score metrics.

The results showed that ResNet excelled with 86% accuracy, slightly higher than ConvMixer which recorded 84% accuracy. ResNet's residual structure overcomes the problem of vanishing gradients, while ConvMixer offers better training speed. Both models showed high performance, although the challenge of similar patterns between classes was still an obstacle. As a result, ResNet provides better results in detecting gastrointestinal diseases, but ConvMixer is also a promising alternative. Further development with more diverse datasets is needed to improve model performance.