Optimization of deep learning-based denoising for arterial spin labeling: Effects of averaging and training...

Systematic study of the effects of averaging and other relevant training strategies in deep learning (DL)-based denoising is required to optimize such processing pipelines for improving the quality of arterial spin labeling (ASL) images.

Different averaging strategies, including windowed and interleaved averaging methods, and different levels of averaging before and after convolutional neural network-based and transformer-based denoising were studied. The experiments were performed on 152 single-delay ASL scans from 152 subjects, including pulsed and pseudo-continuous ASL acquisitions. Four-fold cross-validation was implemented in all experiments. The effect of including calibration scans (M₀) was studied and compared across images of different levels of signal-to-noise ratio (SNR). The generalizability of DL denoising was examined in experiments using low-SNR ground truth in training. The results were assessed using image-quality metrics including structural similarity, peak SNR, and normalized mean absolute error.

Including M₀ was almost always beneficial, with a dependence on the SNR of the input ASL images. Windowed averaging outperformed interleaved averaging, supporting the practice of reducing scan time. Averaging of ASL images before DL denoising was more advantageous than averaging after. Matching the SNR levels of the images in training and inferencing was important for optimal performance. These findings were consistent across convolutional neural network-based and transformer-based models. The generalizability of DL-based denoising was confirmed, and its capability to reduce artifacts was observed.

This study supports the use of DL-based denoising in improving the image quality of ASL and reducing scan time and provides insights to help optimize DL-denoising pipelines.