Benefits of spatial uncertainty aggregation for segmentation in digital pathology

Area of Science:

• Medical Imaging Analysis
• Deep Learning
• Computational Pathology

Background:

• Uncertainty estimation in deep learning (DL) is crucial for medical applications and domain shift challenges.
• Integrating uncertainty estimation effectively into DL systems remains a complex task.
• False negative (FN) predictions in segmentation tasks can have significant clinical implications.

Purpose of the Study:

• To evaluate the Spatial Uncertainty Aggregation (SUA) framework for enhancing uncertainty estimation in DL segmentation.
• To determine if SUA improves the correlation between uncertainty estimates and FN predictions.
• To assess if SUA translates to tangible improvements in segmentation performance.

Main Methods:

• The SUA framework processes negative prediction regions to detect FNs using aggregated uncertainty scores.
• SUA can augment existing uncertainty estimation methods.
• Comparison with a baseline approach that processes individual pixel uncertainties.

Main Results:

• SUA successfully identifies FN regions.
• SUA achieved a metric of 0.92 (in-domain) and 0.85 (domain-shift), significantly outperforming the baseline (0.81 and 0.48).
• SUA demonstrated superior general segmentation performance compared to the baseline uncertainty method.

Conclusions:

• The proposed SUA framework effectively incorporates and utilizes uncertainty estimates for FN detection in DL segmentation for histopathology.
• SUA offers significant advantages over independent pixel uncertainty assessment.
• The framework shows promise for improving the reliability of DL segmentation in medical applications.