Introduction
Diffusion-Weighted MRI (DWI) is a powerful tool for calculating the Apparent Diffusion Coefficient (ADC), a quantitative biomarker used to distinguish between benign and malignant tumors. However, DWI is notoriously sensitive to motion. In preclinical studies, researchers often use radial k-space sampling (Rad-DW-SE) because it is inherently motion-insensitive.
The primary drawback of radial sampling is its slow acquisition speed; a standard high-quality scan can take 25 minutes. While undersampling can accelerate this process, it typically results in noisy images and inaccurate ADC maps. This paper presents DeepADC-Net, a hybrid CNN-Transformer model designed to recover high-quality ADC maps from highly undersampled data, achieving up to 8x acceleration.
Methodology
DeepADC-Net addresses the limitations of standard deep learning approaches through three key architectural and procedural choices:
- Hybrid CNN-Transformer Architecture: While Convolutional Neural Networks (CNNs) excel at local feature extraction, they often lack global context. DeepADC-Net integrates a Vision Transformer into the network's bottleneck, utilizing a self-attention mechanism to capture global spatial relationships across the image.
- Multi-Channel Physics Input: The model processes accelerated DW images at multiple b-values alongside an initial, coarse ADC map calculated via standard curve-fitting.
- Physics-Informed Loss Functions: The model is regularized by the monoexponential diffusion model. The training objective optimizes three complementary losses: the difference between predicted and ground-truth ADC maps, the baseline ($S_0$) signal difference, and a "DWI loss" ensuring the predicted ADC map can mathematically reconstruct the original DW images.
The inclusion of the Bottleneck Transformer is particularly effective at mitigating streaking artifacts common in radial undersampling by leveraging global context.
Experimental Results
The model was trained and validated using a dataset of 183 genetically engineered mice (pancreatic cancer model) on a 9.4T scanner. DeepADC-Net was benchmarked against simulated 4x and 8x undersampled data, as well as real-world 4x accelerated scans.
- Quantitative Superiority: At 4x acceleration, DeepADC-Net outperformed state-of-the-art methods (e.g., U-Net, DenseU-Net) by 3.15% in Correlation Coefficient and 1.29 dB in PSNR.
- Robustness at 8x Acceleration: Even at extreme 8x acceleration (reducing 403 radial views to 50), the model maintained a high structural similarity (SSIM) of 0.9937, whereas standard least-squares fitting dropped to 0.8486.
- Regional Accuracy: The model maintained high fidelity in specific regions of interest (ROIs) such as tumors, kidneys, and muscles, ensuring clinical relevance for diagnostic tasks.
Discussion and Conclusion
Ablation studies confirmed that physics regularization via the monoexponential model fitting term significantly improved ADC map accuracy compared to standard image-to-image translation. Furthermore, directly computing the ADC map within the deep learning process proved more robust to varying signal-to-noise ratios than indirect estimation methods.
While the current study focused on preclinical mouse models at 9.4T, future work will explore the model's performance on clinical scanners and more complex sampling patterns. The authors have made their code and dataset publicly available to support further research in accelerated MRI.
Summary updated