1271 - Distortion Correction In Diffusion MRI for Radiotherapy Planning: From Neuro to Pelvic Applications

10:50am - 10:55am ET

Room 160

Presenter(s)

Deepak Khuntia, MD, FASTRO - Varian Medical Systems/Precision Cancer Specialists, Palo Alto, California

C. Eichner¹, S. Qiu², R. Miron³, Y. Li², N. Janardhanan², B. Clifford⁴, M. Mostapha², M. Nadar², O. Darwish⁵, T. Feiweier⁵, M. Schneider⁶, and D. Khuntia⁷; ¹Siemens Healthineers AG, Forchcheim, Germany, ²Siemens Healthineers, Princeton, NJ, ³Siemens Industry Software Romania, Brasow, Romania, ⁴Siemens Medical Solutions USA, Inc., Boston, MA, ⁵Siemens Healthineers AG, Erlangen, Germany, ⁶Siemens Healthineers, Erlangen, Bayern, Germany, ⁷East Bay Radiation Oncology Center/Eden Medical Center, Castro Valley, CA

Purpose/Objective(s):

Diffusion MRI (dMRI) is a promising functional imaging modality in radiation therapy (RT). Its sensitivity to tumor cellularity, quantitative nature, and non-invasiveness make it well suited for tumor detection, target definition, response assessment, and biologically guided treatment adaptation. However, dMRI relies on EPI acquisitions that are susceptible to patient specific B0-induced distortions, potentially leading to misregistration, contouring inaccuracies, and compromised quantitative parameters - major limitations for RT workflows. Reversed phase-encoding (PE) based iterative distortion correction is available as a research option in neuroimaging, but its long processing times, parameter sensitivity, and limited robustness outside the brain hinder clinical use in RT. We here develop a fast, generalized deep-learning (DL) distortion-correction method suitable for multiple RT-relevant anatomies and compatible with routine clinical workflows.

Materials/Methods:

A multi-anatomy dataset of brain, head/neck, liver, and pelvis was acquired from healthy subjects in diagnostic and RT treatment position on Siemens Healthineers MRI systems - yielding ~85,000 dMRI slices. Subject-level splitting produced 2680 training, and 448 testing scans. The proposed DL model uses Conv2Former architecture with deformable convolutions to estimate the B0 field from a reversed PE image pair. Training used a combination of unsupervised and supervised losses. Distortion correction during deployment is performed via unwarping and intensity compensation.

Results:

The DL approach showed robust performance across all anatomies. Compared to conventional iterative correction (e.g., FSL TopUp), the network achieved improved quantitative metrics (NMSE 0.042 vs. 0.048; PSNR 33.04 dB vs. 32.45 dB; SSIM 0.883 vs. 0.877). Qualitatively, the DL model showed greater robustness in regions with severe susceptibility effects, avoiding blurring seen with iterative methods. Runtime improved considerably: ~0.12 s/vol for DL vs. ~24.2 min/vol for TopUp, enabling real-time correction. A prototype on a 3T Vida RT MRI system demonstrated feasibility of scanner-integrated distortion correction.

Conclusion:

We present a fast and practical DL-based method to correct geometric distortions in dMRI across anatomies commonly treated in RT. The approach requires only a short reference acquisition (~5s) and generates corrected images directly during the reconstruction pipeline, without additional workflow steps. Accurate geometry in dMRI is essential in RT, as distortions can affect target definition, margin selection, and confidence in functional imaging used for response assessment. Providing distortion-corrected dMRI at the scanner improves anatomical fidelity in soft-tissue regions and facilitates integration of dMRI into simulation and adaptive treatment workflows. Overall, this method reduces key barriers to clinical adoption and enables more reliable use of dMRI in RT.