Self-Supervised Deep Learning Framework for Rician Distribution Based Denoising and Modeling of Multi-b Prostate...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Diffusion-Weighted MRI

Background:

• Diffusion-weighted (DW) imaging is crucial for characterizing tissue microstructure.
• Traditional denoising and bias correction methods in DW imaging can be computationally intensive and may introduce artifacts.
• Accurate modeling of the diffusion signal is essential for reliable biomarker quantification.

Purpose of the Study:

• To evaluate convolutional neural networks (CNNs) for enhanced and accelerated denoising and Rician bias correction in multi-b value DW images.
• To assess the performance of different CNN architectures and signal models for simultaneous signal modeling and bias correction.
• To compare the CNN-based approach with a conventional model-based method (OBSIDIAN) in terms of image quality, parameter estimation, and computation time.

Main Methods:

• Self-supervised training of CNNs (U-Net, Attention U-Net, Residual Attention U-Net) using multi-b value prostate DW images from 46 individuals.
• CNNs were trained to output model parameter maps for synthesizing DW images and ADC maps, incorporating Rician bias correction.
• Exploration of signal models including biexponential, kurtosis, and gamma distributions, with and without Rician bias correction, and with/without noise map input.

Main Results:

• CNN-generated synthetic DW images demonstrated comparable quality to the OBSIDIAN method.
• CNN models produced less noisy ADC and parameter maps than OBSIDIAN, with good quantitative agreement for ADC values.
• Rician bias correction was essential for accurate results, while noise map input had a less pronounced effect.
• Computation time was reduced from several hours (OBSIDIAN) to seconds using CNNs.

Conclusions:

• CNN-based methods offer a promising approach for clinical DW imaging, delivering higher quality images and biomarker maps.
• The significant reduction in computation time makes CNNs highly attractive for practical applications.
• Further research with larger datasets is recommended to enhance generalizability and robustness of the CNN models.