Clustering-Guided Twin Contrastive Learning for Endomicroscopy Image Classification

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Learning discriminative features for medical image analysis is challenging due to limited annotated datasets.
• Developing effective representation learning methods is crucial for computer-aided diagnosis.

Purpose of the Study:

• To propose a novel clustering-guided twin-contrastive learning framework (CTCL) for learning discriminative representations from limited-scale, unlabeled probe-based confocal laser endomicroscopy (pCLE) images.
• To improve intra-class tightness and inter-class variability for more informative representations in gastrointestinal (GI) tumor classification.

Main Methods:

• Developed a clustering-guided twin-contrastive learning framework (CTCL) that aligns semantically related and class-consistent samples.
• Incorporated geometric invariance and noise tolerance by treating rotated and differently noised pCLE images as the same instance.
• Optimized CTCL using an end-to-end expectation-maximization framework.

Main Results:

• CTCL-based representations achieved competitive performance, robustness, and transferability compared to state-of-the-art methods.
• Achieved 75.60%/78.45% and 64.12%/77.37% top-1 accuracy on linear evaluation and few-shot classification tasks, respectively, outperforming previous best results.
• Demonstrated improved intra-class tightness and inter-class variability in the learned embedding space.

Conclusions:

• CTCL effectively learns discriminative representations from limited pCLE data for GI tumor classification.
• The proposed method offers a promising approach for automated, fast, and high-precision diagnosis of GI tumors.
• CTCL has the potential to aid pathologists in determining tumor development stages using pCLE images.