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

Convolutional neural networks (CNNs) are evaluated for improved and accelerated denoising and Rician bias correction in multi-b DW images with simultaneous signal modeling.

Prostate diffusion images from 46 individuals acquired at 20 linearly distributed b-values ( $b_{\max }=2000s/{\text{mm}}^2)$ without averaging were used for self-supervised CNN training. CNNs were trained to output model parameter maps, from which DW images and ADC maps were synthesized, and to account for Rician bias. CNN architectures included were: conventional U-Net, Attention U-Net, and Residual Attention U-Net with biexponential, kurtosis, and gamma distribution as signal models. Moreover, approaches without Rician bias correction and with noise maps as additional input were explored.

For all signal models, synthetic DW images generated from CNN output are of similar quality compared to OBSIDIAN, a model-based method employing iterative pixel-wise fitting. Computed ADC and parameter maps obtained with CNN models were less noisy than those obtained with OBSIDIAN. Good quantitative agreement with OBSIDIAN in ROI-averaged ADC values and, to a slightly lesser extent, kurtosis model parameters was observed. Results with omitted Rician bias correction clearly deviate, while the impact of the noise map input for the CNNs is less pronounced. Computation time was reduced to seconds compared to several hours for OBSIDIAN.

The CNN-based method has potential in clinical use for achieving higher quality DW images and biomarker maps with drastically reduced computation time. Further refinement and larger training datasets are needed to achieve better generalizability and increased robustness.