Empirical Study of Attention-Based Models for Automatic Classification of Gastrointestinal Endoscopy Images

Automatic and accurate analysis of medical images is a subject of great importance in our current society. In particular, this work focuses on gastrointestinal endoscopy images, as the study of these images helps to detect possible health conditions in those regions. Published works on this topic mainly used traditional classification methods (e.g., Support VectorMachines) or more modern techniques, such as Convolutional Neural Networks. However, little attention has been paid to more recent approaches such as Transformers or, in general, Attention-based Deep Neural Networks. This work aims to evaluate the performance of state-of-the-art attention-based models on the problem of classification of gastrointestinal endoscopy images. The experimental results on the challenging Hyper-Kvasir dataset indicate that attention-based models achieve performance equal to or better than that obtained by previous models, needing fewer parameters. In addition, a new state of the art on Hyper-Kvasir (i.e., 0.636 F1-Macro) is obtained by the fusion of two MobileViT models with only 20M parameters. The source code will be published here: https://github.com/richardesp/Attention-based-models-for-Hyper-Kvasir/.