Clinical Scorecard: A Two-Phase Deep Learning Approach for Detecting Laterally Spreading Tumors Utilizing Self-Supervised Learning and Few-Shot Classification Techniques
At a Glance
Category
Detail
Condition
Laterally spreading tumors (LSTs), precursor lesions of colorectal cancer
Key Mechanisms
Two-stage AI framework using DINO-based self-supervised pretraining and Prototypical Networks for few-shot classification
Target Population
Patients undergoing colonoscopy, aged ≥18 years
Care Setting
Colonoscopy and endoscopy clinical settings
Key Highlights
LSTs are lesions ≥10 mm growing horizontally along the intestinal wall with high malignant potential and difficult detection.
Proposed AI framework reduces reliance on large annotated datasets by leveraging self-supervised learning and few-shot classification.
Achieved clinically relevant detection performance using only 2799 labeled training images, validated with rigorous multi-expert annotation.
Guideline-Based Recommendations
Diagnosis
Diagnosis of LSTs currently relies on visual inspection by experienced endoscopists but is subject to false negatives.
Use internationally recognized diagnostic criteria for LSTs as per Kudo et al.
Employ multi-tier expert consensus annotation for accurate LST classification.
Management
Early detection of LSTs is critical to reduce progression to high-grade intraepithelial neoplasia and colorectal cancer.
Integrate AI-assisted detection tools to support endoscopists and improve diagnostic accuracy.
Monitoring & Follow-up
Maintain quality control of colonoscopy images with adequate bowel preparation and image clarity.
Use stratified dataset splits and cross-validation to monitor AI model performance and generalizability.
Risks
False negatives due to operator experience and visual fatigue in manual detection.
Potential overfitting and suboptimal performance of AI models trained on limited labeled data without self-supervised pretraining.
Patient & Prescribing Data
12,376 colonoscopy cases with 3.02% diagnosed with LSTs, ages 15–94 years
AI model trained on 2799 labeled images with additional 150,168 unlabeled images for pretraining; balanced datasets used to improve detection accuracy.
Clinical Best Practices
Ensure high-quality colonoscopy images with adequate bowel preparation (Boston Bowel Preparation Scale ≥6) and proper illumination.
Adopt rigorous multi-expert annotation protocols for dataset labeling to establish reliable ground truth.
Utilize self-supervised learning on large unlabeled datasets to mitigate data scarcity in rare lesion detection.
Apply few-shot classification techniques to enable effective learning from limited labeled examples.
Validate AI models with held-out test sets and cross-validation to ensure robustness and clinical applicability.
