Exploring end-to-end learning for high-resolution 3D Spin Echo MRI

• Applicant: Professor Dr. Moritz Zaiss
• Subject Area: Medical Physics, Biomedical Technology
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 558090876

Turbo Spin Echo (TSE) sequences are pivotal in clinical MRI, particularly for diagnosing diseases such as cancer, multiple sclerosis, and vascular disorders. However, current TSE techniques face limitations that hinder their diagnostic performance, especially for fast sequences with long echo trains. Despite advancements like segmentation, parallel imaging, deep learning, and flip angle train optimization, these limitations remain partly unresolved on the image acquisition side. In this project, we aim to gain new knowledge of optimal TSE sequence design through end-to-end learning, encompassing TSE flip angle and k-space sampling design, advanced reconstruction, and simultaneously trained image processing neural networks. My group has recently developed MR-zero, a framework that treats the entire MRI process as a differentiable computational chain, allowing for end-to-end optimization. This approach enables tailored sequence customization, thereby optimizing image quality for specific diagnostic needs. Our goal in this project is to gain new insights through end-to-end learning-based sequence design for optimal TSE image acquisition, allowing us to advance image quality, sharpness, and flexibility in contrast. The fundamental knowledge gained will guide 2D and 3D TSE design in general at any field strength or target region. Specifically, we aim to determine the optimal flip angle train design for minimal blurring in 3D-TSE and to understand the optimal signal formation process to be SNR-, PSF-, and SAR-optimal. We also intend to explore joint end-to-end optimization of flip angle train and sampling pattern, along with subsequent image processing networks for 3D-TSE, to understand their interaction and achieve simultaneous optimality. Finally, we aim to demonstrate the benefits of new 3D-TSE in brain and knee MRI at 3T and 7T. The target 3D-TSE is projected to achieve whole-brain coverage in under 4 minutes, with a 0.66 mm isotropic resolution. We expect it to exceed the contrast, quality, and sharpness of current approaches, as quantified by RMSE, SSIM, and PSNR metric of high-segmented TSEs. While our initial focus is on 3T and 7T applications in knee and brain imaging, this approach has broader implications and can be adapted for other field strengths such as 1.5T and 0.55T, and other target regions, providing general new insights into TSE that are also valid for 2D-TSE.

• DFG Programme: Research Grants