TLDR: A study evaluated DeepFoqus-Accelerate, an FDA-cleared deep learning algorithm, for accelerating brain MRI scans. It found that the algorithm reliably enables up to fourfold acceleration, reducing scan times by 75%, while fully preserving diagnostic image quality. Both expert radiologists and quantitative metrics confirmed the AI-reconstructed images were clinically equivalent to standard scans, promising improved patient comfort and workflow efficiency in neuroimaging.
Magnetic Resonance Imaging (MRI) is a vital diagnostic tool, offering detailed views of soft tissues, crucial for identifying neurological, vascular, and musculoskeletal conditions. However, the lengthy scan times associated with MRI have long been a significant drawback. These extended durations can lead to patient discomfort, especially for children or those with claustrophobia, and increase the risk of motion artifacts, which can degrade image quality and necessitate repeat scans. Furthermore, long scan times reduce the number of patients that can be scanned, creating bottlenecks and increasing operational costs for healthcare providers.
To tackle these challenges, researchers are developing accelerated MRI acquisition techniques. While traditional methods like parallel imaging and compressed sensing have been implemented, they typically offer limited acceleration (up to twofold) and can sometimes compromise image quality or introduce artifacts at higher speeds. Recent advancements in machine learning, particularly deep learning (DL), are now showing great promise in overcoming these limitations by enabling high-quality MRI reconstruction from heavily undersampled data.
A new study, titled “Evaluation of Machine Learning Reconstruction Techniques for Accelerated Brain MRI Scans,” investigates the performance of a deep learning-based MRI reconstruction algorithm called DeepFoqus-Accelerate. This algorithm, developed by Foqus Technologies Inc., is FDA-cleared (510(k) K241982) and designed to significantly speed up brain MRI scans while maintaining diagnostic quality. The research, conducted by Jonathan I. Mandel, Shivaprakash Hiremath, Hedyeh Keshtgar, Timothy Scholl, and Sadegh Raeisi, aimed to rigorously validate DeepFoqus-Accelerate’s capabilities using both public and prospective clinical data. You can read the full research paper here: Research Paper.
How the Study Was Conducted
The study employed a mixed retrospective-prospective design. Retrospective data came from the publicly available fastMRI dataset, which includes raw k-space brain MRI data from various Siemens platforms (1.5T and 3T) and a wide range of clinical pathologies. For prospective data, 18 healthy adult volunteers were recruited between January 2024 and March 2025. These participants underwent both standard-of-care (SOC) and fourfold accelerated brain MRI scans on a 3T GE Discovery MR750 system. The accelerations were simulated by retrospectively reducing phase-encoding steps to achieve 2x, 3x, and 4x undersampling, effectively reducing scan time by up to 75%.
DeepFoqus-Accelerate version 1.1 was used to reconstruct the undersampled k-space data. This algorithm utilizes a proprietary deep neural network architecture, trained on a large, diverse MRI dataset separate from the study population to minimize bias.
Expert Review and Quantitative Analysis
To assess image quality, five experienced raters—three board-certified neuroradiologists and two MRI technologists—independently reviewed 36 paired datasets (SOC and AI-reconstructed accelerated scans). They scored the overall image quality on a 5-point Likert scale, where 1 meant non-diagnostic and 5 meant identical to SOC, with a focus on diagnostic utility and artifact presence.
In addition to expert review, a quantitative assessment was performed on 408 scans, yielding 1224 sets. Objective image similarity metrics were used, including Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Haar wavelet-based Perceptual Similarity Index (HaarPSI). These metrics are known to correlate strongly with how radiologists perceive image quality.
Key Findings
The results were highly encouraging. In the qualitative evaluations, none of the AI-reconstructed images were rated below a diagnostic quality threshold (all scores were 3 or higher), and a remarkable 95% of the images scored 4 or higher, indicating high diagnostic confidence. The mean score across all expert readers was 4.38 out of 5. While some minor inter-reader variability was observed, rare artifacts present in the reconstructed images did not impede lesion detection or anatomical delineation.
The quantitative evaluations further supported these findings, showing high structural similarity between DeepFoqus-Accelerate reconstructions and SOC images across all sequences (T1, T2, and FLAIR). The mean SSIM was 0.959 ± 0.034, PSNR averaged above 41 dB, and HaarPSI values exceeded 0.94. More than 90% of the AI reconstructions demonstrated an SSIM greater than 0.90, consistently indicating high similarity to the reference images.
Implications for Clinical Practice
This study demonstrates that DeepFoqus-Accelerate can reliably enable up to fourfold accelerated brain MRI acquisition without compromising diagnostic image quality. The ability to reduce scan times by up to 75% translates into significant benefits for patients and healthcare systems. Patients will experience less discomfort, a decreased risk of motion artifacts, and potentially fewer sedation requirements. For clinics, it means improved access to MRI services and greater patient throughput, leading to more efficient operations.
The findings also suggest that these AI-reconstructed images are suitable for advanced quantitative tasks, such as automated volumetric analysis and lesion segmentation, further expanding the clinical utility of the technique without sacrificing critical diagnostic features.
Also Read:
- Enhancing Medical Image AI with HU-Based Foreground Masking
- XBusNet: A New Multimodal Model for Precise Breast Ultrasound Segmentation
Future Directions
Despite the rigorous assessments, the study acknowledges some limitations, including a relatively small, single-center prospective cohort and the absence of prospective multi-vendor scanner diversity. Future research will aim for multi-center and multi-vendor designs to further validate and expand the use of AI-accelerated MRI in routine practice across a wider range of clinical indications.
In conclusion, the DeepFoqus-Accelerate algorithm represents a significant step forward in neuroimaging. By maintaining diagnostic quality while drastically reducing scan times, it promises to enhance patient experience, improve workflow efficiency, and advance clinical practice in MRI.
