Journal Article

Mitigating the Impact of MR Sequence Parameters: Increasing the Robustness of DL‐Based Cortical Thickness Estimates

McKinley, Richard, Wiest, Roland, Capiglioni, Milena, Romascano, David, Salmen, Anke, Rebsamen, Michael, Hoepner, Robert, Rummel, Christian, Radojewski, Piotr, Blattner, Timo, Pistor, Maximilian

2026

Overview

Cortical thickness measurements from MRI are increasingly used as biomarkers for neurodegenerative disease progression. However, variations in MRI acquisition parameters, such as inversion time (TI) and repetition time (TR), which are common in clinical settings, can compromise the reliability and sensitivity of these measurements. We fine‐tuned a deep‐learning‐based segmentation tool (DL+DiReCT) to reduce its dependence to image contrast variations by training it on simulated MPRAGE images derived from quantitative relaxation maps. Fine‐tuning markedly reduced contrast sensitivity, with the Pearson correlation coefficient decreasing from −0.644 -0.644 to 0.094 0.094 . Evaluation on a synthetic atrophy dataset demonstrated that our model accurately replicated atrophy trends with minimal underestimation, outperforming FreeSurfer and SynthSeg. When applied to a dataset of relapsing–remitting multiple sclerosis (RRMS) patients, the fine‐tuned model showed a substantial reduction in contrast sensitivity and maintained stable performance after controlling for covariates such as age, sex, field strength, and Expanded Disability Status Scale (EDSS) score. Overall, the proposed approach achieves robust contrast invariance without sacrificing sensitivity to cortical atrophy, offering a practical improvement for longitudinal and multi‐center clinical studies. Key Points Changes in MR acquisition settings, specifically of TI and TR, which are common in the clinical setting, affect WM/GM contrast, which in turn affects cortical thickness measurements. Deep learning models can be fine‐tuned on synthetic MRI simulation across different contrasts for more robust cortical thickness measurements. Left: Our finetuned model reduces dependence on image contrast significantly (top), while conserving sensitivity to synthetic atrophy (bottom). Right: For the finetuning procedure, MRIs with varying grey‐white contrast were synthesized using Deichmann's equations on T1 and PD maps. The segmentation ground truth for all synthetic images was generated by running FreeSurfer on the MRI with highest image contrast. Cortical thickness estimated with various methods was finally compared.

• Publisher: John Wiley & Sons, Inc,Wiley
• Subject: Adult / Atrophy / Atrophy - pathology / Bias / Biomarkers / Brain Cortical Thickness / brain morphometry / Cerebral Cortex - diagnostic imaging / Cerebral Cortex - pathology / contrast / Correlation coefficient / Correlation coefficients / cortical thickness / Datasets / Deep Learning / Disease progression / Female / Field strength / Humans / Image contrast / Image Processing, Computer-Assisted - methods / Image Processing, Computer-Assisted - standards / Magnetic resonance imaging / Magnetic Resonance Imaging - methods / Magnetic Resonance Imaging - standards / Male / Medical imaging / Middle Aged / Multiple sclerosis / Multiple Sclerosis, Relapsing-Remitting - diagnostic imaging / Multiple Sclerosis, Relapsing-Remitting - pathology / Neurodegenerative diseases / Neuroimaging - methods / Neuroimaging - standards / Parameter robustness / robustness / Scanners / Sensitivity / Thickness measurement