Subclass-Aware Contrastive Semi-Supervised Learning for Inflammatory Bowel Disease Classification from Colonoscopy...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate classification of Inflammatory Bowel Disease (IBD), encompassing Crohn's disease (CD) and ulcerative colitis (UC), from colonoscopy images is crucial for patient diagnosis and treatment.
• Deep learning models for IBD classification require substantial labeled data, which is often scarce, posing a significant challenge.
• Existing semi-supervised learning methods using pseudo-labeling struggle with IBD images due to high intra-class variability and subtle inter-class differences, leading to inaccurate pseudo-labels and confirmation bias.

Purpose of the Study:

• To develop an advanced semi-supervised learning method, Subclass-Aware Contrastive Semi-Supervised Learning (SACSSL), for accurate IBD classification from colonoscopy images.
• To overcome the limitations of traditional pseudo-labeling methods by mitigating confirmation bias and handling intra-class heterogeneity.
• To enhance the performance of deep learning models for IBD classification using limited labeled data.

Main Methods:

• Proposed SACSSL method integrates a subclass-aware contrastive module into a pseudo-labeling framework (e.g., FixMatch).
• Uncertain unlabeled samples are processed with instance-level contrastive loss to reduce confirmation bias.
• Confident unlabeled samples are clustered into fine-grained subclasses using prototypes and supervised contrastive loss to improve inter-class separability and intra-class diversity.

Main Results:

• SACSSL achieved state-of-the-art performance on both an in-house IBD classification dataset (Daping) and a public UC severity grading dataset (LIMUC) in a semi-supervised setting.
• With only 20% labeled data, SACSSL reached 93.2% accuracy and 80.1% F1-score on the Daping dataset, closely approaching the fully supervised performance.
• On the LIMUC dataset, SACSSL achieved 76.4% accuracy and 68.9% F1-score, demonstrating its effectiveness across different IBD-related tasks.

Conclusions:

• The proposed SACSSL method effectively enhances semi-supervised colonoscopy image classification for IBD.
• SACSSL successfully addresses challenges of inaccurate pseudo-labels, confirmation bias, and intra-class variability in IBD image analysis.
• The method demonstrates significant potential for improving IBD diagnosis and treatment stratification with limited labeled data.