Uncertainty Estimation and Out-of-Distribution Detection for Deep Learning-Based Image Reconstruction Using the Local...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image reconstruction is crucial for diagnostics.
• Supervised deep learning models struggle with unseen data distributions.
• Existing uncertainty estimation methods do not assess training distribution fit.

Purpose of the Study:

• To develop a method for distinguishing in-distribution from out-of-distribution images in deep learning medical imaging.
• To assess the reliability of deep learning models when encountering novel data.
• To improve the diagnostic accuracy of reconstructed medical images.

Main Methods:

• Proposed a method based on the local Lipschitz metric to detect out-of-distribution images.
• Validated the method using the AUTOMAP architecture for Magnetic Resonance Imaging (MRI) reconstruction.
• Expanded validation to MRI denoising and Computed Tomography (CT) reconstruction using UNET architectures.

Main Results:

• Achieved 99.94% Area Under the Curve (AUC) for distinguishing in-distribution from out-of-distribution images.
• Demonstrated a strong correlation (Spearman's rho = 0.8475) between local Lipschitz values and Mean Absolute Error (MAE).
• Outperformed baseline methods like Monte-Carlo dropout, deep ensembles, and Mean Variance Estimation.

Conclusions:

• The local Lipschitz metric effectively identifies out-of-distribution data in medical imaging.
• The correlation between local Lipschitz values and MAE can guide data augmentation and uncertainty reduction.
• The proposed method is versatile, applicable to various architectures and medical imaging tasks, enhancing diagnostic reliability.