Imaging Transformer for MRI Denoising: a Scalable Model Architecture that enables SNR ≪ 1 Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) is crucial for cardiovascular assessment, but image quality is often compromised by low signal-to-noise ratios (SNR).
• Denoising techniques are essential to enhance MRI image quality for reliable clinical interpretation and quantitative analysis.
• Existing denoising methods struggle to maintain image fidelity and anatomical detail, particularly at very low SNR levels.

Purpose of the Study:

• To introduce a flexible and scalable Imaging Transformer (IT) architecture for multi-dimensional medical imaging data.
• To apply the proposed IT architecture to Magnetic Resonance Imaging (MRI) denoising, focusing on scenarios with very low input SNR.
• To evaluate the IT model's performance against established convolutional and transformer-based denoising methods.

Main Methods:

• Developed three attention modules: spatial local, spatial global, and frame attention, integrated into an attention-cell-block design for 5D tensor processing.
• Built a High-Resolution Network (HRNet) backbone to host the IT blocks, processing 2D, 2D+T, and 3D imaging data.
• Trained and tested IT models on a large dataset (206,677 training, 7,267 testing cine series) across ten SNR levels (0.05-8.0), comparing against seven baseline models and assessing scalability with models up to 218 million parameters.

Main Results:

• IT models significantly outperformed all baseline models across tested SNR levels, with the most substantial performance gains observed at low SNR.
• The IT-218m model achieved the highest Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR), preserving image quality and anatomical details even at SNR 0.2.
• Clinical experts confirmed that IT model outputs at SNR 0.2 or higher provided the same interpretation as ground-truth images, with accurate ejection fraction (EF) measurements.

Conclusions:

• The Imaging Transformer (IT) model demonstrates robust performance, scalability, and versatility for MRI denoising applications.
• The IT model effectively restores image quality, enabling confident clinical reading and precise EF measurements, even in challenging low-SNR conditions (down to SNR 0.2).
• This architecture offers a promising solution for improving the diagnostic utility of MRI in various clinical settings.