Magnetic Resonance Imaging (MRI) is a widely used diagnostic tool that offers detailed visualisation of internal body structures. However, the process is often time-consuming, presenting economic and practical challenges. Efforts to improve the speed and efficiency of MRI scans have led to the use of parallel imaging (PI) and other acceleration techniques. Still, these can introduce noise and image artefacts, compromising image quality. A recent study published in Radiology Advances proposes a novel deep-learning-based denoising technique designed to enhance the quality of brain MRIs acquired using accelerated imaging protocols.
Deep Learning for MRI Denoising
Traditional MRI acquisition techniques can be slow, leading to longer scan times and patient discomfort. While PI techniques have helped reduce scan times, they often introduce noise and artefacts, particularly at high acceleration factors. Unlike other approaches that require access to raw data, the proposed deep-learning-based denoising method operates entirely in the image domain, making it straightforward to implement across different scanner models and vendors.
The deep learning model is a U-Net architecture that enhances the quality of MRI images by removing noise and correcting residual errors that PI algorithms fail to address. This model was trained on retrospective data, simulating various image degradations, and tested on both retrospective and prospective data to ensure robustness. Prospective testing, in particular, demonstrated the method’s generalisability across different clinical environments and scanner types, further validating the model’s applicability.
Evaluation and Key Findings
The study evaluated the model's performance both qualitatively and quantitatively. Five board-certified radiologists assessed the images qualitatively based on criteria such as image quality, noise, artefact reduction, and visibility of anatomical features. Quantitative analysis focused on signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and spatial resolution.
- Qualitative Improvements: The model-enhanced images showed significant improvements in overall image quality, noise reduction, and artefact removal compared to the original accelerated images. The visibility of both normal anatomical structures and pathological features improved as well. The enhanced images were rated as non-inferior or even superior to baseline images (lower acceleration scans), maintaining the necessary diagnostic quality.
- Quantitative Enhancements: The AI-enhanced images demonstrated notable increases in SNR and CNR across all types of sequences (sagittal T1-weighted, axial T2-weighted, and axial fluid-attenuated inversion recovery (FLAIR)). Improved SNR and CNR indicate a clearer distinction between structures and better overall image clarity. Spatial resolution, measured by the full width at half maximum (FWHM), was preserved or improved, indicating that the model did not introduce blurring or reduce the sharpness of the images.
- Reduced Scan Time: One of the main benefits of this deep-learning model is the reduction in scan time. The model enables shorter MRI sequences without compromising image quality by allowing higher acceleration factors (3 or 4, depending on the vendor). On average, the total scan time was reduced by approximately 29%, with sequence-specific reductions of 19-41% depending on the scanner model and protocol. These findings highlight the potential for faster imaging protocols to improve patient comfort, reduce operational costs, and increase scanner throughput.
Implications and Future Directions
The deep-learning-based denoising model presents a significant advancement in MRI imaging, addressing critical challenges associated with high acceleration factors. By enhancing image quality and enabling faster scan times, this technique holds promise for both clinical and research settings. Faster scans could lead to improved patient experiences, reduced healthcare costs, and the ability to serve a larger number of patients without sacrificing diagnostic accuracy.
However, the study has some limitations, including its focus on a specific set of sequences and a single field strength (3 Tesla). Future research could explore the model's applicability to other anatomical regions, MRI orientations, contrasts, and field strengths (e.g., 1.5 Tesla), potentially broadening its clinical utility.
The use of deep learning in enhancing MRI imaging offers a promising avenue for improving the efficiency and quality of brain scans. The AI-based denoising technique evaluated in this study effectively reduced noise and artefacts while preserving or enhancing diagnostic image quality. With scan times reduced by nearly a third, the potential impact on clinical workflows and patient experiences is substantial. Integrating AI-driven enhancements is supposed to play a pivotal role in advancing diagnostic imaging.
Source Credit: Radiology Advances
Image Credit: iStock
