Real-time, inline quantitative MRI enabled by scanner-integrated machine learning: a proof of principle with NODDI

Purpose: The clinical feasibility and translation of many advanced quantitative MRI (qMRI) techniques are inhibited by their restriction to ‘research mode’, due to resource-intensive, offline parameter estimation. This work aimed to achieve ‘clinical mode’ qMRI, by real-time, inline parameter estimation with a trained neural network (NN) fully integrated into a vendor’s image reconstruction environment, therefore facilitating and encouraging clinical adoption of advanced qMRI techniques. Methods: The Siemens Image Calculation Environment (ICE) pipeline was customised to deploy trained NNs for advanced diffusion MRI parameter estimation with Open Neural Network Exchange (ONNX) Runtime. Two fully-connected NNs were trained offline with data synthesised with the neurite orientation dispersion and density imaging (NODDI) model, using either conventionally estimated (NNMLE) or ground truth (NNGT) parameters as training labels. The strategy was demonstrated online in two healthy volunteers (one rescanned) and evaluated offline with synthetic data, testing two diffusion protocols. Results: NNs were successfully integrated and deployed natively in ICE, performing inline, whole-brain, in vivo NODDI parameter estimation in <10 seconds. The proposed workflow was reproducible across protocols, volunteers and rescans. DICOM parametric maps were exported from the scanner for further analyses. Comparisons between NNMLE and NNGT suggested NNMLE parameter estimates to be more consistent with conventional fitting, a finding supported by offline evaluations. Conclusion: Real-time, inline parameter estimation with the proposed generalisable framework resolves a key practical barrier to the potential clinical uptake of advanced qMRI methods, enabling their efficient integration into clinical workflows. Next steps include incorporation of pre-processing methods and evaluation in pathology.

💡 Research Summary

**
Quantitative magnetic resonance imaging (qMRI) offers biologically meaningful biomarkers by fitting biophysical models to MR signals, but its clinical adoption has been hampered by the need for time‑consuming, offline parameter estimation. This paper demonstrates a complete solution that brings advanced qMRI, specifically Neurite Orientation Dispersion and Density Imaging (NODDI), into the clinical workflow by performing real‑time, inline parameter estimation directly on the scanner.

The authors customized Siemens’ Image Calculation Environment (ICE) to embed the Open Neural Network Exchange (ONNX) Runtime. After the standard image reconstruction pipeline assembles a 4‑D diffusion dataset, the ICE program normalizes each voxel’s signal and feeds it to a pre‑trained fully‑connected neural network (NN) that instantly outputs the three NODDI parameters: orientation dispersion index (ODI), neurite density index (NDI), and free‑water fraction (FWF). The resulting parametric maps are written as a separate DICOM series, allowing immediate visualization on the console and automatic transfer to the hospital’s PACS and reporting systems.

Two NNs were trained offline using synthetic diffusion data generated from the NODDI forward model. Training data covered two diffusion protocols (a 2‑shell scheme with 81 encodings and a 3‑shell scheme with 116 encodings) and a realistic range of signal‑to‑noise ratios. For each synthetic voxel, ground‑truth (GT) parameter values were drawn from uniform distributions, a random fiber orientation was assigned, and Rician‑distributed noise was added. The authors created two training label strategies: (1) conventional maximum‑likelihood estimator (MLE) fits of the synthetic signals (NN‑MLE) and (2) the original GT parameters (NN‑GT). Both networks share the same architecture (four hidden layers of 256 units each) and were implemented in PyTorch before being exported to ONNX.

In vivo validation was performed on a 3 T MAGNETOM Vida scanner with two healthy volunteers (ages 31 and 25). Each subject was scanned with both diffusion protocols; one subject was rescanned without repositioning, yielding six datasets. After loading the ONNX files onto the scanner console, the custom ICE program reconstructed the raw data and performed voxel‑wise NN inference. Whole‑brain NODDI maps were generated in under 10 seconds, compared with conventional MLE fits that required roughly 3 minutes per slice (≈2 hours for the whole brain) on a research workstation.

Quantitative comparison used tissue masks (white matter, gray matter, CSF) derived from SynthSeg. Scatter plots and mean absolute differences (MAD) showed that NN‑MLE produced parameter estimates closely matching the conventional MLE, especially for NDI and FWF in white matter and ODI and FWF in gray matter. NN‑GT, while still producing plausible maps, exhibited larger bias and variance, particularly at higher free‑water fractions. Synthetic test data (100 000 voxels) were also evaluated: bias vectors for NN‑MLE overlapped with those of MLE, confirming that the NN inherits the intrinsic limitations of the NODDI model rather than introducing additional errors. NN‑GT showed systematic over‑ or under‑estimation in certain regions, leading to higher MAD values. The three‑shell protocol reduced overall bias and narrowed the performance gap between NN‑MLE and NN‑GT.

The study highlights several practical implications. By embedding inference in the scanner’s native reconstruction pipeline, no external GPU servers, third‑party frameworks (e.g., Gadgetron), or manual data export/import steps are required. This dramatically lowers the hardware and IT expertise barrier for clinical sites. Immediate availability of quantitative maps enables potential applications in acute care (e.g., stroke assessment) and longitudinal monitoring where rapid feedback is essential.

Limitations include the current absence of preprocessing steps such as eddy‑current correction, motion correction, and Gibbs ringing removal; the presented workflow assumes high‑quality data. Moreover, the NNs are trained for specific diffusion encoding schemes and SNR ranges, so substantial protocol changes would necessitate retraining. Future work will focus on integrating preprocessing modules into ICE, extending validation to pathological cohorts (e.g., multiple sclerosis, brain tumors), and generalizing the framework to other advanced models such as CHARMED or MAP‑MRI.

In conclusion, the authors provide a proof‑of‑concept that real‑time, scanner‑integrated qMRI is feasible. Their approach eliminates the major practical bottleneck—offline, computationally intensive fitting—thereby paving the way for routine clinical use of sophisticated diffusion models and potentially transforming neuroimaging diagnostics and research.