Gastrointestinal diseases diagnosis using capsule endoscopy and YOLOv8.

Abstract

Gastrointestinal (GI) diseases represent a substantial global health concern that annually affect millions of individuals and result in nearly two million deaths, highlighting the urgent need for early and accurate diagnosis, as undiagnosed cases can be life-threatening. Wireless Capsule Endoscopy (WCE) is a cutting-edge technology that enables the visualization of gastrointestinal diseases. By capturing thousands of frames per patient, it reduces the risk of human error and increases the accuracy of diagnoses. The enormous volume of images generated by WCE poses a significant challenge for manual diagnosis, prompting the development of computer-aided techniques to enhance the diagnostic process with high accuracy and within a short period. Moreover, deep learning algorithms have demonstrated remarkable performance in medical imaging tasks and especially in GI diseases classification, leveragin gth evast amounts of data generated by WCE to improve diagnostic precision and enhance patient outcomes. This project represents a technique aimed at developing a robust YOLOv8-cls model for gastrointestinal disease classification. By training the model on the Kvasir dataset, the backbone of YOLOv8-cls learns to capture robust and informative features from the input images. These features serve as a powerful representation of the image content. Finally, the extracted features are fed into a classification head, whic hi sfine-tuned to predict the class of the input WCE image, enabling accurate diagnosis of gastrointestinal diseases. In our project we conducted a serie of four experiments to develop a high-performing YOLOv8-cls model for gastrointestinal disease classification. The initial experiment identified the best-performin gYOLOv8-cl svariant, which was then optimized usin ghyperparameter tuning. The final model wa sconstructed using the mixture of experts technique, combining the best-performing variant with optimized hyperparameters. The top-performing variant from experiment 1 achieved an accuracy of 92.7%, but exhibited confusion between specifi cclasses. Hyperparameter tuning and the mixture of experts approach improved the model’s performance. Our proposed model achieved a testing accuracy of 96.25%, precision of 96.28%, and recall of 96.25%, with a 4% increase in testing accuracy, precision, and recall compared to the initial model.