Meta-UNet: enhancing skin-lesion segmentation with multimodal feature integration and uncertainty estimation

Area of Science:

• Medical image analysis
• Machine learning in healthcare
• Computational pathology

Background:

• Accurate medical image segmentation is vital for diagnostics.
• Standard segmentation models like U-Net can have limitations in accuracy and uncertainty estimation.
• Lesion-specific metadata offers potential for improving segmentation performance.

Purpose of the Study:

• To investigate the impact of integrating lesion-specific metadata into the U-Net architecture for enhanced medical image segmentation.
• To evaluate different metadata integration strategies and their effect on segmentation accuracy.
• To assess the reduction in predictive uncertainty achieved by metadata integration.

Main Methods:

• Modified the U-Net architecture to incorporate metadata (Meta-UNet).
• Evaluated integration strategies: addition, weighted addition, and embedding layers.
• Developed Bayesian Meta-UNet with Monte Carlo Dropout (MCD) for uncertainty quantification using Confidence Maps, Entropy, Mutual Information, and Expected Pairwise Kullback-Leibler divergence (EPKL).
• Introduced an aggregation strategy for a comprehensive image-level uncertainty score.

Main Results:

• Meta-UNet demonstrated superior performance over standard U-Net on PH2, ISIC 2018, and HAM10000 datasets.
• Achieved higher accuracy and Intersection over Union (IoU) metrics across all tested datasets.
• Demonstrated a reduction in predictive uncertainty, indicating increased model confidence with metadata integration.

Conclusions:

• Integrating lesion-specific metadata into U-Net architecture significantly enhances segmentation accuracy.
• Metadata integration effectively reduces predictive uncertainty, leading to more reliable segmentation outcomes.
• The proposed Meta-UNet approach holds promise for strengthening diagnostic segmentation pipelines in clinical practice.