Uncertainty-Aware Training for Ophthalmic Segmentation Using MedSAM.

Area of Science:

• Ophthalmic imaging analysis
• Medical deep learning
• Uncertainty quantification

Background:

• Deep learning (DL) models require enhanced interpretability and learning guidance.
• Uncertainty quantification (UQ) is a key area for improving DL model reliability.
• Ophthalmic segmentation tasks benefit from precise pixel-level classification.

Purpose of the Study:

• Introduce Uncertainty-Aware Training (UAT) to augment DL loss functions with uncertainty maps.
• Enhance DL model performance and interpretability by focusing on areas of high uncertainty.
• Improve accuracy in ophthalmic segmentation tasks like geographic atrophy (GA), optic cup (OC), and foveal avascular zone (FAZ).

Main Methods:

• Applied UAT to three ophthalmic segmentation tasks: GA, OC, and FAZ.
• Weighted binary cross-entropy loss function using uncertainty maps to focus on ambiguous regions.
• Experimented with entropy-based UQ and conformal prediction techniques.
• Evaluated UAT on a fine-tuned state-of-the-art foundational model.

Main Results:

• Entropy-weighted maps at loss calculation consistently improved performance across all datasets.
• Conformal prediction (Least Ambiguous Set-Valued Classifier) enhanced GA and OC segmentation.
• UAT integration was feasible without significant modifications to training structures.

Conclusions:

• UAT effectively enhances DL model performance and interpretability by incorporating uncertainty alongside error.
• The method improves segmentation accuracy in ophthalmic applications.
• UAT's lightweight integration facilitates practical adoption and improved model performance.