DermaCalibra: A Robust and Explainable Multimodal Framework for Skin Lesion Diagnosis via Bayesian Uncertainty and...

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate skin lesion diagnosis is challenged by class imbalance and visual similarity.
• Existing multimodal learning methods often neglect predictive uncertainty and clinical metadata integration.
• Small-scale, imbalanced datasets pose significant hurdles for deep learning models in dermatology.

Purpose of the Study:

• To introduce DermaCalibra, a robust, explainable multimodal framework for skin lesion diagnosis.
• To address class imbalance and integrate heterogeneous clinical metadata effectively.
• To improve diagnostic accuracy and reliability, especially in resource-limited settings.

Main Methods:

• Developed the Attention-Based Multimodal Channel Recalibration (AMCR) module with Bayesian uncertainty estimation and focal loss adjustment.
• Implemented the Metadata-Driven Dynamic Feature Modulation and Cross-Attention Fusion (MDFM-CAF) module for resolving visual ambiguity using clinical context.
• Utilized the Gradient Feature Attribution (GFA) module for pixel-level diagnostic heatmaps and metadata importance.

Main Results:

• Achieved 84.2% balanced accuracy (BACC) and 96.9% Macro AUC on the PAD-UFES-20 dataset, outperforming SOTA by 3.6% BACC.
• Demonstrated robust generalizability across diverse clinical settings via external validation on unseen hospital and synthetic datasets without retraining.
• Validated the framework's ability to prioritize features from underrepresented classes and resolve inter-class visual ambiguity.

Conclusions:

• DermaCalibra bridges deep learning complexity with clinical intuition via uncertainty-aware reasoning and interpretability.
• The framework offers a reliable and scalable computer-aided diagnostic tool for early skin lesion detection.
• DermaCalibra shows significant potential for improving dermatological diagnostics in resource-limited environments.