Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep convolutional neural networks (CNNs) excel in medical image segmentation but often lack uncertainty estimation.
• Existing methods struggle to differentiate between model (epistemic) and image-based (aleatoric) uncertainties.

Purpose of the Study:

• To analyze and quantify epistemic and aleatoric uncertainties in CNN-based 2D and 3D medical image segmentation.
• To propose a novel test-time augmentation (TTA) method for estimating aleatoric uncertainty.
• To develop a theoretical framework for TTA in medical image segmentation.

Main Methods:

• Proposed a TTA-based aleatoric uncertainty estimation by analyzing segmentation output variations under image transformations.
• Developed a theoretical formulation for TTA using Monte Carlo simulation and an image acquisition model.
• Compared and combined TTA-based aleatoric uncertainty with dropout-based model uncertainty.

Main Results:

• TTA-based aleatoric uncertainty provided superior uncertainty estimation compared to dropout-based model uncertainty alone.
• The proposed method effectively reduced overconfident incorrect segmentation predictions.
• TTA outperformed both single-prediction baselines and dropout-based multiple predictions in segmentation tasks.

Conclusions:

• Test-time augmentation offers a robust framework for quantifying aleatoric uncertainty in medical image segmentation.
• The proposed TTA method enhances segmentation reliability by providing better uncertainty estimates and reducing errors.
• This work advances the understanding and application of uncertainty quantification in deep learning for medical imaging.