Ensemble learning via supervision augmentation for white matter hyperintensity segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning for Segmentation
• Computational Neuroscience

Background:

• White matter hyperintensity (WMH) segmentation is challenging due to ambiguous lesion boundaries and inter-observer variability, leading to noisy annotations.
• Deep neural networks (DNNs) can overfit noisy labels, resulting in biased models with poor generalization, especially when multiple annotations per image are unavailable.
• Existing methods often rely on multiple annotations, which are not always feasible in real-world clinical settings.

Purpose of the Study:

• To propose a novel supervision augmentation (SA) method combined with ensemble learning (SA-EN) to enhance the generalization ability of WMH segmentation models.
• To address the challenge of noisy and uncertain annotations in medical image segmentation when only single annotations are available.
• To improve the accuracy and robustness of deep learning models for WMH segmentation.

Main Methods:

• Developed a supervision augmentation (SA) method to generate diverse supervision signals by estimating annotation uncertainty from single, ambiguous annotations.
• Integrated SA with ensemble learning (EN), training different base learners with varied supervision information derived from SA.
• Evaluated the SA-EN method on two WMH segmentation datasets, comparing its performance against state-of-the-art ensemble techniques.

Main Results:

• SA-EN achieved optimal accuracy compared to other ensemble methods on WMH segmentation tasks.
• The proposed method demonstrated superior effectiveness on small datasets, making it suitable for medical image segmentation scenarios with limited annotations.
• SA-EN successfully captured both aleatoric (annotation uncertainty) and epistemic (model uncertainty) types of uncertainty.

Conclusions:

• The SA-EN approach effectively improves the generalization ability and accuracy of deep learning models for white matter hyperintensity segmentation.
• This method offers a robust solution for medical image segmentation challenges characterized by noisy annotations and limited data.
• SA-EN provides a valuable framework for uncertainty quantification in deep learning models applied to medical imaging.