Improving Endoscopy Lesion Classification Using Self-Supervised Deep Learning.

Area of Science:

• Medical image analysis
• Machine learning in healthcare
• Computational pathology

Background:

• Gastritis atrophy (GA) and intestinal metaplasia (IM) are precancerous gastric lesions.
• Early detection of GA and IM is critical for preventing gastric cancer progression.
• Current detection methods benefit from advanced machine learning techniques.

Purpose of the Study:

• To evaluate the effectiveness of self-supervised learning (SSL) for detecting gastritis atrophy (GA) and intestinal metaplasia (IM).
• To compare SSL performance against supervised learning baselines across varying data availability.
• To investigate the influence of data augmentation on SSL for gastric lesion detection.

Main Methods:

• Experiments were conducted using the Chengdu dataset.
• SSL models were trained and evaluated with different proportions of annotated data.
• Performance was compared to a supervised learning baseline using classification accuracy.
• The impact of data augmentation techniques in contrastive learning was analyzed.

Main Results:

• SSL achieved classification accuracy comparable to supervised learning when using all available annotated data (81.52% vs 81.76%).
• In low-data regimes (12.5% annotated data), SSL demonstrated an accuracy improvement of approximately 1.5% over supervised learning (73.00% vs 71.52%).
• SSL performance was found to be sensitive to the choice of data augmentation strategies.

Conclusions:

• SSL models show significant potential in leveraging unlabeled data for improved performance and generalization in gastric lesion detection.
• SSL offers a robust alternative to supervised learning, particularly in data-scarce medical imaging scenarios.
• Further research into optimized data augmentation frameworks for SSL in gastric lesion detection is warranted.