Using an artificial neural network for fast mapping of the oxygen extraction fraction with combined QSM and quantitative BOLD

Abstract

Purpose To apply an artificial neural network (ANN) for fast and robust quantification of the oxygen extraction fraction (OEF) from a combined QSM and quantitative BOLD analysis of gradient echo data and to compare the ANN to a traditional quasi‐Newton (QN) method for numerical optimization. Methods Random combinations of OEF, deoxygenated blood volume (ν), R2, and nonblood magnetic susceptibility (χnb) with each parameter following a Gaussian distribution that represented physiological gray matter and white matter values were used to simulate quantitative BOLD signals and QSM values. An ANN was trained with the simulated data with added Gaussian noise. The ANN was applied to multigradient echo brain data of 7 healthy subjects, and the reconstructed parameters and maps were compared to QN results using Student t test and Bland‐Altman analysis. Results Intersubject means and SDs of gray matter were OEF =43.5±0.8%, R2=13.5±0.3 Hz, ν=3.4±0.1%, χnb=-25±5 ppb for ANN; and OEF = 43.8±5.2%, R2=12.2±0.8 Hz, ν=4.2±0.6%, χnb=-39±7 ppb for QN, with a significant difference (P<0.05) for R2, ν, and χnb. For white matter, they were OEF = 47.5±1.1%, R2=17.1±0.4 Hz, ν=2.5±0.2%, χnb=-38±5 ppb for ANN; and OEF =42.3±5.6%, R2=16.7±0.7 Hz, ν=2.9±0.3%, χnb=-45±9 ppb for QN, with a significant difference (P<0.05) for OEF and ν. ANN revealed more gray–white matter contrast but less intersubject variation in OEF than QN. In contrast to QN, the ANN reconstruction did not need an additional sequence for parameter initialization and took approximately 1 s rather than roughly 1 h. Conclusion ANNs allow faster and, with regard to initialization, more robust reconstruction of OEF maps with lower intersubject variation than QN approaches.

Quantification of brain oxygen extraction and metabolism with [15O]-gas PET: A technical review in the era of PET/MRI

“…Despite its relative accessibility and flexibility, OEF measurement by MRI is often hindered by complex contributions of multiple physiological parameters and low deoxyhemoglobin signal, as with quantitative BOLD MRI. A recently proposed solution to these limitations is use of an artificial neural network to emulate the curve-fitting of dynamic quantitative BOLD signals and estimate OEF ( Hubertus et al, 2019 ). This image transformation network takes the multi-echo gradient echo signals and magnetic susceptibility maps as inputs and outputs a final OEF map.…”

“…The neural network provided smaller inter-subject variation in OEF that was more in line with PET literature, compared to standard quasi-Newton fitting of quantitative BOLD. However, the network training was performed purely on numerical datasets with simulated noise ( Hubertus et al, 2019 ), which does not realistically model all acquisition and physiological features of OEF images. If paired oxygenation-sensitive MRI and [ 15 O]-gas PET signals are available, particularly from a simultaneous PET/MRI acquisition, a similar network can be trained using PET as the “gold standard” to enhance OEF performance.…”

Consensus-based technical recommendations for clinical translation of renal BOLD MRI

“…There is a future role for artificial neural networks/ artificial intelligence (AI) in solving cortex and medulla segmentation problems on BOLD MRI data acquired in patients with renal disease. To date, there are no published studies on AI in renal BOLD MRI, but artificial neural networks have been used in the brain for quantitative analysis of BOLD MRI and quantitative susceptibility mapping data [49].…”

Deep learning using a biophysical model for robust and accelerated reconstruction of quantitative, artifact‐free and denoised images

“…The major limitation of the traditional learning paradigm is that it requires large amounts of ground‐truth data for training, which can be challenging to obtain in practice. One strategy, adopted by, 11,12 is to synthetically generate such data in individual voxels using the signal model (eg, Equation ). However, this strategy can be sensitive to the mismatch between the actual statistical distribution of entries in (which might have spatial statistical dependencies) and the assumed distribution used in simulation (often independent in space).…”

“…The key benefit of using deep learning over the traditional voxel‐by‐voxel analysis lies in its ability to deal with noisy input and superior computational speed. This was recently demonstrated on a related problem of estimating the oxygen extraction fraction (OEF) maps, where an ANN, with a single hidden layer, was trained and applied voxel‐wise to produce the desired quantitative parameters given the mGRE signal input 11,12 . The ground‐truth datasets for supervised training in 11,12 were generated using simulated signals based on the analytical model 13,14 …”