Two-Phase Deep Learning for Laterally Spreading Tumor Detection Using Self-Supervised and Few-Shot Learning
Overview
This study presents a novel two-stage AI framework combining self-supervised learning and few-shot classification to detect laterally spreading tumors (LSTs) in colonoscopy images. Utilizing 150,168 unlabeled images for pretraining and only 2799 labeled images for training, the approach achieves clinically relevant performance while addressing data scarcity and annotation challenges.
Background
Laterally spreading tumors (LSTs) are flat colorectal lesions ≥10 mm in diameter that pose a high risk for progression to colorectal cancer. Their subtle morphology makes detection challenging, often leading to false negatives in endoscopic diagnosis. Traditional deep learning methods require large annotated datasets, which are difficult to obtain for LSTs due to their low prevalence and annotation costs. Integrating self-supervised learning with few-shot classification offers a promising solution to improve detection accuracy with limited labeled data.
Data Highlights
Dataset
Number of Images
Purpose
Unlabeled Dataset
150,168
Self-supervised pretraining
Labeled Dataset
4,000
Training, validation, and testing
Training Set
2,799
Model training
Validation Set
600
Hyperparameter tuning and model selection
Test Set
601
Final model evaluation
Key Findings
The proposed two-stage framework leverages DINO-based self-supervised pretraining on a large unlabeled colonoscopy image set to learn robust feature representations.
Prototypical Networks were adapted for few-shot classification, enabling effective LST detection from a limited labeled dataset of 2799 images.
The model maintained clinically relevant diagnostic performance despite the significant reduction in labeled training data compared to conventional supervised methods.
Comprehensive evaluations, including feature space analysis and decision curve analysis, demonstrated model robustness and potential for real-time clinical application.
Stratified dataset splitting ensured balanced LST-to-normal ratios, supporting unbiased model training and evaluation.
Clinical Implications
This approach reduces dependency on large annotated datasets, addressing a major bottleneck in LST detection. The integration of self-supervised and few-shot learning techniques can enhance early identification of LSTs during colonoscopy, potentially improving patient outcomes by facilitating timely intervention. Additionally, the model's real-time inference capability supports practical deployment in clinical endoscopy workflows.
Conclusion
The study successfully demonstrates that combining self-supervised pretraining with few-shot classification enables accurate and efficient detection of laterally spreading tumors using limited labeled data. This framework holds promise for improving colorectal cancer screening through enhanced AI-assisted endoscopic diagnosis.
References
Kudo et al. -- Diagnostic criteria for LSTs
Li et al. -- Deep learning in lumbar MRI diagnosis
Shanghai General Hospital Ethics Committee (20251124011538817)
