Improving image classification of gastrointestinal endoscopy using curriculum self-supervised learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Endoscopic diagnosis of gastrointestinal (GI) disorders faces challenges due to non-specific symptoms and limited accessibility.
• Supervised machine learning aids diagnosis but is hindered by a scarcity of expert-labeled GI images.
• Existing methods struggle with the vast amount of unlabeled endoscopic data.

Purpose of the Study:

• To develop a novel curriculum self-supervised learning framework for enhanced GI disorder diagnosis.
• To leverage large unlabeled datasets for pre-training diagnostic models.
• To improve the accuracy and efficiency of diagnosing GI disorders using endoscopic imagery.

Main Methods:

• A curriculum self-supervised learning framework inspired by human learning was proposed.
• The HyperKvasir dataset, containing 100k unlabeled and 10k labeled GI images, was utilized.
• The framework involved pre-training on unlabeled data and fine-tuning on labeled data.

Main Results:

• The proposed method achieved a top-1 accuracy of 88.92% and an F1 score of 73.39%.
• This represents a 2.1% increase in top-1 accuracy and a 1.9% increase in F1 score compared to vanilla SimSiam.
• The results demonstrate the effectiveness of curriculum self-supervised learning in this domain.

Conclusions:

• Curriculum self-supervised learning effectively utilizes unlabeled GI images for improved disorder diagnosis.
• The framework offers a promising approach for more accurate and efficient GI endoscopy.
• This study paves the way for advanced AI-driven diagnostic tools in gastroenterology.