Related Experiment Videos
Vision transformer-based uncertainty quantification for triaging skin lesions: a probabilistic framework for
Jafaridarabjerdi Mahin1,2, Lin Li2,3
1Faculty of Medicine, Dalian University of Technology, Dalian, Liaoning, China.
Frontiers in Bioengineering and Biotechnology
|July 14, 2026
Summary
This study introduces a new AI model for melanoma detection, improving diagnostic accuracy by combining visual and clinical data. The system quantifies uncertainty, leading to safer biopsy suggestions and reduced diagnostic errors in dermatology.
Area of Science:
- Artificial Intelligence in Dermatology
- Medical Image Analysis
- Machine Learning for Healthcare
Background:
- Deep learning models for skin lesion detection show promise but lack clinical integration due to overconfidence and reliance on visual data alone.
- Existing systems often fail to incorporate patient history, leading to potential diagnostic errors in melanoma triage.
Purpose of the Study:
- To develop a novel probabilistic model for enhanced skin lesion analysis and biopsy suggestion.
- To improve the safety and reliability of automated diagnostic systems in clinical practice through uncertainty quantification.
Main Methods:
- Utilized a Vision Transformer architecture (Swin Transformer) for complex feature extraction from dermoscopic images.
- Implemented a cross-attention system to integrate multimodal data, correlating visual features with patient clinical data (age, gender, lesion location).
- Incorporated a Dirichlet distribution in the decision layer for uncertainty estimation and biopsy need prediction.
Main Results:
- Achieved 92.398% accuracy and 0.924 AUROC in classifying skin lesions.
- Reduced Expected Calibration Error (ECE) by 0.031 through evidential learning, improving model calibration.
- Demonstrated 92.37% sensitivity for detecting lesions requiring biopsy and identified 89.5% of ambiguous images for expert review.
Conclusions:
- The multimodal, uncertainty-aware system enhances diagnostic accuracy and reduces errors in dermatology triage by integrating clinical context.
- This AI framework offers a computational safety layer, paving the way for Trustworthy AI in clinical healthcare workflows.