Denoising and uncertainty estimation in parameter mapping with approximate Bayesian deep image priors.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Machine Learning

Background:

• Parameter mapping in medical imaging often suffers from noise and high uncertainty.
• Conventional methods struggle to effectively denoise these parameter maps.

Purpose of the Study:

• To address noisy parameter maps and high uncertainties by framing parameter mapping as a denoising task using Deep Image Priors (DIP).
• To extend the denoising capabilities of DIP to tissue parameter map generation.

Main Methods:

• Utilized Deep Image Prior (DIP) by treating an image-generating network's output as a parametrization of tissue parameter maps.
• Employed an untrained convolutional neural network (CNN) for implicit denoising, filtering low-level image features.
• Integrated uncertainty estimation using Monte Carlo (MC) dropout for voxel-wise uncertainty quantification.
• Developed a modular approach allowing adaptation to various applications like T1 mapping, T2 mapping, and apparent diffusion coefficient mapping.

Main Results:

• Demonstrated successful adaptation of DIP across multiple parameter mapping applications.
• Achieved significant noise reduction and decreased uncertainty in parameter maps compared to conventional techniques.
• Identified extended computational time and potential bias introduction from the denoising prior as limitations.

Conclusions:

• Deep Image Prior (DIP) effectively denoises parameter mapping and is applicable across various scenarios with minimal tuning.
• The implementation's ease, stemming from the absence of training data requirements, is a key advantage.
• While computationally intensive, MC dropout-derived uncertainty information enhances robustness and provides valuable insights when calibrated.