DeepFoqus-Accelerate MRI Reconstruction

• The paper demonstrates that DeepFoqus-Accelerate maintains diagnostic image quality at 4× acceleration, with 95% of reconstructions scoring ≥4 on expert ratings.
• It employs a proprietary deep neural network on undersampled k-space data to reconstruct both 2D and 3D T1, T2, and FLAIR brain MRI with high SSIM, PSNR, and HaarPSI metrics.
• The study suggests practical benefits including up to 75% reduction in scan time, improved patient comfort, and enhanced workflow efficiency, despite limited prospective validation.

Searching arXiv for the cited paper and topic context. DeepFoqus-Accelerate is a deep learning MRI reconstruction algorithm for undersampled brain MRI, described as an FDA-cleared (510(k) K241982), k-space–based deep learning reconstruction system designed to reconstruct diagnostically usable images from phase-encoding–undersampled acquisitions, enabling up to fourfold acceleration of brain MRI (Mandel et al., 8 Sep 2025). In the reported evaluation, it was applied to accelerated 2D and 3D T1-, T2-, and FLAIR-weighted brain MRI and compared against standard-of-care (SOC) fully sampled reference images. The central clinical motivation is the long acquisition time of conventional brain MRI, which is associated with patient discomfort, motion, repeat imaging, sedation needs in some populations, scheduling bottlenecks, and cost; within that context, DeepFoqus-Accelerate is presented as a reconstruction-stage method for substantial scan-time reduction without sacrificing diagnostic quality (Mandel et al., 8 Sep 2025).

1. System definition and reconstruction setting

DeepFoqus-Accelerate is evaluated in the setting of accelerated brain MRI reconstruction from phase-encoding–undersampled acquisitions. In the study, undersampling was applied in the phase-encoding direction, and the algorithm reconstructed 2D and 3D T1-, T2-, and FLAIR-weighted sequences for comparison with SOC fully sampled images (Mandel et al., 8 Sep 2025). The paper specifies that SOC denotes the standard-of-care fully sampled MRI acquisition/reconstruction used as the clinical reference standard.

The algorithm is described at a high level rather than at the architectural level. The study used DeepFoqus-Accelerate v1.1, characterized as a proprietary deep neural network architecture trained on a large heterogeneous MRI dataset external to the study population (Mandel et al., 8 Sep 2025). The paper provides that it is k-space based and operates on undersampled k-space data in the reconstruction pipeline, rather than as a simple post-processing denoiser on SOC images. At the same time, the paper does not disclose the exact network architecture, number of cascades or layers, whether data consistency blocks are used, loss functions, optimizer or training schedule, exact training set composition, inference runtime, vendor adaptation strategy, or raw k-space coil combination details (Mandel et al., 8 Sep 2025). This suggests that the paper is primarily an evaluation study rather than a methodological disclosure of the reconstruction model internals.

2. Study design, cohorts, and acquisition protocol

The evaluation was a mixed retrospective–prospective study (Mandel et al., 8 Sep 2025). The retrospective component used the public fastMRI brain dataset, described as containing multi-coil raw k-space brain MRI from 1.5T and 3T Siemens scanners, with a range of ages, sexes, and clinical pathologies. The fastMRI portion included fully sampled 2D axial T1, T2, and FLAIR sequences (Mandel et al., 8 Sep 2025).

The prospective component included 18 healthy adult volunteers, comprising 8 men and 10 women, with mean age 35.5 years and range 22–66 years, recruited and scanned between January 2024 and March 2025 (Mandel et al., 8 Sep 2025). Prospective scans were acquired on a 3T GE Discovery MR750.

The prospective protocol included both 2D and 3D brain MRI sequences. The 2D sequences were axial, coronal, and sagittal T1 spin echo; axial, coronal, and sagittal T2 fast spin echo; and axial, coronal, and sagittal T2 FLAIR. The 3D sequences were axial T1 BRAVO, sagittal T2, and coronal FLAIR (Mandel et al., 8 Sep 2025).

Acceleration was simulated retrospectively by reducing phase-encoding lines to produce 2×, 3×, and 4× acceleration, corresponding approximately to 50%, 66%, and 75% scan-time reduction (Mandel et al., 8 Sep 2025). The supplemental methods specify an equispaced phase-encoding mask with an 8% fully sampled k-space center fraction. For 4× acceleration, only 25% of phase-encoding lines are preserved overall, while all central 8% of k-space lines are retained (Mandel et al., 8 Sep 2025).

Two evaluation sets were defined. For qualitative reader review, 36 paired datasets were selected. For quantitative analysis, 408 scans with multiple acceleration rates generated 1224 evaluated datasets (Mandel et al., 8 Sep 2025).

3. Comparative methodology and evaluation metrics

The comparison framework proceeded from fully sampled data, retrospectively undersampled the phase-encoding dimension at 2×, 3×, or 4×, reconstructed using DeepFoqus-Accelerate, and compared the reconstructed images with SOC using expert reader scoring and quantitative similarity metrics (Mandel et al., 8 Sep 2025). The central tested claim was whether AI reconstruction remained diagnostically equivalent to SOC, especially at 4× acceleration.

Three quantitative image-similarity metrics were used across the 408 scans and 1224 datasets: Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Haar wavelet-based Perceptual Similarity Index (HaarPSI) (Mandel et al., 8 Sep 2025). The paper provides the standard SSIM definition:

$$\mathrm{SSIM}(x,y)=\frac{(2\mu_x\mu_y+c_1)(2\sigma_{xy}+c_2)}{(\mu_x^2+\mu_y^2+c_1)(\sigma_x^2+\sigma_y^2+c_2)}$$

where $\mu_x,\mu_y$ are means, $\sigma_x^2,\sigma_y^2$ variances, and $\sigma_{xy}$ covariance (Mandel et al., 8 Sep 2025). For PSNR, the reported formula is:

$$\mathrm{PSNR}=10\log_{10}\left(\frac{\mathrm{MAX}_I^2}{\mathrm{MSE}}\right)$$

where $\mathrm{MAX}_I$ is the maximum possible pixel intensity and MSE is mean squared error (Mandel et al., 8 Sep 2025). HaarPSI is described as a perceptual image-quality metric based on Haar wavelet features, intended to reflect human-perceived similarity better than purely pixelwise error measures; no explicit formula is given in the paper (Mandel et al., 8 Sep 2025).

For the reader study, five independent readers participated: 3 board-certified neuroradiologists with 10–15 years of experience and 2 MRI technologists with 10 and 18 years of experience (Mandel et al., 8 Sep 2025). They were independent of the software development team. Readers scored overall image quality relative to SOC using a 5-point Likert scale focused on diagnostic utility and artifact presence, with 1 indicating non-diagnostic, 2 severe distortion or artifact or noise affecting interpretation, 3 minimum acceptable diagnostic quality, 4 almost perfect reconstruction with only minor differences, and 5 diagnostically identical to original or SOC (Mandel et al., 8 Sep 2025). Diagnostic acceptability was therefore defined as score $\ge 3$.

Inter-rater agreement was assessed using weighted Cohen’s kappa, and scoring pattern consistency was also assessed with Spearman correlation (Mandel et al., 8 Sep 2025). The study reports a significance threshold of $p < 0.05$ for Spearman correlations, with some pairwise correlations significant at $p < 0.01$.

4. Qualitative and quantitative performance

The main reported finding is that DeepFoqus-Accelerate preserved diagnostic image quality even at 4× acceleration (Mandel et al., 8 Sep 2025). Across all expert readers, no AI-reconstructed scan scored below 3, corresponding to 0% judged non-diagnostic, and 95% scored $\ge 4$ (Mandel et al., 8 Sep 2025). The mean Likert score was $\mu_x,\mu_y$0, and the median score was 4.4. Given the scale definition, these values place nearly all evaluated reconstructions in the range between minimum acceptable diagnostic quality and almost perfect reconstruction, with most in the latter range.

Quantitative similarity metrics were likewise high. Across the full quantitative dataset, SSIM was $\mu_x,\mu_y$1 with 95% CI $\mu_x,\mu_y$2–$\mu_x,\mu_y$3, PSNR was $\mu_x,\mu_y$4 dB with 95% CI $\mu_x,\mu_y$5–$\mu_x,\mu_y$6, and HaarPSI was $\mu_x,\mu_y$7 with 95% CI $\mu_x,\mu_y$8–$\mu_x,\mu_y$9 (Mandel et al., 8 Sep 2025). The paper also states that over 90% of AI reconstructions had SSIM $\sigma_x^2,\sigma_y^2$0, PSNR was $\sigma_x^2,\sigma_y^2$1 dB, and HaarPSI was $\sigma_x^2,\sigma_y^2$2, including the abstract formulation that 90% of cases exceeded SSIM 0.90 (Mandel et al., 8 Sep 2025). These results were reported across the 2×, 3×, and 4× acceleration distributions.

For the qualitative-reader subset, the reported values were SSIM $\sigma_x^2,\sigma_y^2$3, PSNR $\sigma_x^2,\sigma_y^2$4 dB, and HaarPSI $\sigma_x^2,\sigma_y^2$5, each with corresponding confidence intervals (Mandel et al., 8 Sep 2025). These are slightly lower than the full-dataset values, which is consistent with the paper’s note that the worst-case qualitative examples also tended to have lower quantitative scores.

Sequence-specific analysis showed variation across contrasts (Mandel et al., 8 Sep 2025):

Sequence	SSIM	PSNR	HaarPSI
T1	$\sigma_x^2,\sigma_y^2$6	$\sigma_x^2,\sigma_y^2$7 dB	$\sigma_x^2,\sigma_y^2$8
T2	$\sigma_x^2,\sigma_y^2$9	$\sigma_{xy}$0 dB	$\sigma_{xy}$1
FLAIR	$\sigma_{xy}$2	$\sigma_{xy}$3 dB	$\sigma_{xy}$4

Sequence-specific results suggest that T2 performed best overall, T1 also performed strongly, and FLAIR was somewhat lower, though still high and still diagnostically acceptable by the reader study (Mandel et al., 8 Sep 2025). Because this contrast dependence is explicitly reported, it is a significant part of the evaluation rather than a secondary detail.

5. Reader agreement, artifacts, and diagnostic interpretation

Inter-rater agreement was reported as slight to moderate, with weighted Cohen’s $\sigma_{xy}$5 ranging from $\sigma_{xy}$6 to $\sigma_{xy}$7 (Mandel et al., 8 Sep 2025). The strongest Spearman concordance was among reviewers 1, 2, and 4, and some pairwise correlations were significant at $\sigma_{xy}$8. The paper emphasizes that reviewer 2 appeared stricter, but this did not change diagnostic conclusions (Mandel et al., 8 Sep 2025).

The study notes rare artifacts, including wrap-around, motion-related distortions, and occasional intensity nonuniformity (Mandel et al., 8 Sep 2025). Importantly, these were said not to impede lesion detection or anatomical delineation. The worst-case examples in the qualitative set had wrap-around and motion artifacts, and such cases also tended to have lower quantitative scores. This addresses a common interpretive concern in AI reconstruction studies: a high average similarity metric does not by itself establish preservation of local diagnostic content. Here, the paper’s claim is narrower and clinically framed: rare artifacts were observed, but they did not affect diagnostic interpretation in the reviewed cases (Mandel et al., 8 Sep 2025).

The paper does not report a formal noninferiority analysis or a hypothesis-test p-value comparing AI versus SOC image-quality distributions (Mandel et al., 8 Sep 2025). Instead, the evidence is based on reader ratings, confidence intervals for similarity metrics, inter-rater statistics, and artifact review. This is important for interpreting the study’s conclusions. The findings support preserved diagnostic quality within the reported evaluation design, but they are not presented in the form of a formal noninferiority statistical framework.

6. Practical implications for neuroimaging workflow

The practical implication emphasized by the authors is up to 75% scan-time reduction from 4× acceleration (Mandel et al., 8 Sep 2025). The supplement gives specific 3D examples: 3D axial T1 BRAVO reduced from 8:15.58 to 2:19.52, 3D sagittal T2 from 14:49.02 to 3:15.90, and 3D coronal FLAIR from 18:28.53 to 5:38.21. These examples operationalize what fourfold acceleration means at the protocol level.

The authors argue that such reductions could yield improved patient comfort, less motion, fewer repeat scans, potentially less sedation, improved scanner throughput, and greater workflow efficiency and access (Mandel et al., 8 Sep 2025). They also claim preservation of image quality sufficient for clinical interpretation and potentially for downstream tasks such as volumetry and lesion segmentation, although those downstream tasks were not directly tested in the study (Mandel et al., 8 Sep 2025). The latter point should therefore be read as an implication rather than as a demonstrated result within the paper.

From a reconstruction-pipeline perspective, the study positions DeepFoqus-Accelerate as a method for accelerating acquisition rather than as a post hoc enhancement of already acquired SOC images (Mandel et al., 8 Sep 2025). That distinction matters in MRI workflow analysis because time savings arise only when undersampling is performed at acquisition and the AI model reconstructs from the resulting k-space data.

7. Limitations, caveats, and evidentiary scope

The paper acknowledges several limitations that bound interpretation of the results (Mandel et al., 8 Sep 2025). The prospective cohort was small, comprising only 18 healthy volunteers, which limits assessment of real-world pathology in the prospective arm. The prospective data were single-center and acquired on a single GE 3T scanner, so external generalizability remains uncertain. Pathology diversity came mainly from fastMRI, whereas the prospective cohort consisted of healthy volunteers rather than a pathology-rich clinical cohort (Mandel et al., 8 Sep 2025).

Reader agreement was only slight to moderate, reflecting subjectivity in image-quality assessment and suggesting caution when interpreting subtle differences (Mandel et al., 8 Sep 2025). Artifacts were rare and reportedly non-limiting, but wrap-around, motion, and intensity nonuniformity remain identified failure modes. Sequence-specific variation is also relevant: FLAIR had the lowest quantitative performance among sequences, and the paper explicitly notes that this may matter because FLAIR is often critical in neuroimaging (Mandel et al., 8 Sep 2025).

Generalizability is limited by the absence of prospective multi-vendor validation beyond the combination of public Siemens data and prospective GE data, by the lack of a multicenter clinical deployment study, and by the absence of an outcome-based evaluation of lesion detection sensitivity and specificity (Mandel et al., 8 Sep 2025). Regulatory status is also distinguished from clinical generalization: although the system is FDA-cleared, the paper still frames broader adoption as requiring further multicenter and multi-vendor validation (Mandel et al., 8 Sep 2025).

Taken together, the study supports DeepFoqus-Accelerate as a clinically plausible reconstruction method for fourfold accelerated brain MRI, with no qualitative score below diagnostic acceptability, 95% of scans rated $\sigma_{xy}$9, and strong similarity to SOC in SSIM, PSNR, and HaarPSI (Mandel et al., 8 Sep 2025). The main caution is that the strongest prospective evidence comes from a small healthy-volunteer cohort, so broader validation in larger, multicenter, pathology-rich populations remains necessary before treating the reported performance as definitive across routine neuroimaging use cases.