cs.CVApr 1, 2026

Automated Detection of Multiple Sclerosis Lesions on 7-tesla MRI Using U-net and Transformer-based Segmentation

Authors: Michael Maynord, Minghui Liu, Cornelia Fermüller, Seongjin Choi, Yuxin Zeng, Shishir Dahal, Daniel M. Harrison

Organizations: Department of Computer, Science University of Maryland College Park, College Park, 20742, MD, USA · Department of Neurology, University of Maryland School of Medicine, Baltimore, MD, 21201, USA · Baltimore VA Medical Center, VA Maryland Healthcare System, Baltimore, MD, 21201, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Ultra-high field 7-tesla (7T) MRI improves visualization of multiple sclerosis (MS) white matter lesions (WML) but differs sufficiently in contrast and artifacts from 1.5-3T imaging - suggesting that widely used automated segmentation tools may not translate directly. We analyzed 7T FLAIR scans and generated reference WML masks from Lesion Segmentation Tool (LST) outputs followed by expert manual revision. As external comparators, we applied LST-LPA and the more recent LST-AI ensemble, both originally developed on lower-field data. We then trained 3D UNETR and SegFormer transformer-based models on 7T FLAIR at multiple resolutions (0.5x0.5x0.5^3, 1.0x1.0x1.0^3, and 1.5x1.5x2.0^3) and evaluated all methods using voxel-wise and lesion-wise metrics from the BraTS 2023 framework. On the held-out test set at native 0.5x0.5x0.5^3 resolution, 7T-trained transformers achieved competitive overlap with LST-AI while recovering additional small lesions that were missed by classical methods, at the cost of some boundary variability and occasional artifact-related false positives. On a held-out 7 T test set, our best transformer model (SegFormer) achieved a voxel-wise Dice of 0.61 and lesion-wise Dice of 0.20, improving on the classical LST-LPA tool (Dice 0.39, lesion-wise Dice 0.02). Performance decreased for models trained on downsampled images, underscoring the value of native 7T resolution for small-lesion detection. By releasing our 7T-trained models, we aim to provide a reproducible, ready-to-use resource for automated lesion quantification in ultra-high field MS research (https://github.com/maynord/7T-MS-lesion-segmentation).

Figures & tables

ModelResolutionTPFPFNSensitivitySpecificity
LST-LPA0.5 ×\times 0.5 ×\times 0.5 mm 325.2748.28.00.82110.9978
LST-AI0.5 ×\times 0.5 ×\times 0.5 mm 329.314.63.90.69800.9997
UNETR0.5 ×\times 0.5 ×\times 0.5 mm 329.1128.54.10.70890.9996
UNETR1.0 ×\times 1.0 ×\times 1.0 mm 328.437.74.80.59530.9995
UNETR1.5 ×\times 1.5 ×\times 2.0 mm 320.065.313.20.36650.9996
SegFormer0.5 ×\times 0.5 ×\times 0.5 mm 327.042.24.60.70720.9997
Table 1: Detection summary across models and resolutions: mean True Positives (TP), False Positives (FP), False Negatives (FN), sensitivity, and specificity on the held-out 7T test set.
ModelResolutionDICELesion-wise DICE
LST-LPA0.5 ×\times 0.5 ×\times 0.5 mm 30.38950.0184
LST-AI0.5 ×\times 0.5 ×\times 0.5 mm 30.68290.4495
UNETR0.5 ×\times 0.5 ×\times 0.5 mm 30.56570.0997
UNETR1.0 ×\times 1.0 ×\times 1.0 mm 30.51800.1534
UNETR1.5 ×\times 1.5 ×\times 2.0 mm 30.34620.0712
SegFormer0.5 ×\times 0.5 ×\times 0.5 mm 30.61200.1967
Table 2: Overlap metrics across models and resolutions: voxel-wise Dice-Sørensen coefficient (DICE) and lesion-wise DICE computed in the native 0.5 mm 3 7T space.
ModelResolutionHD95Lesion-wise HD95
LST-LPA0.5 ×\times 0.5 ×\times 0.5 mm 321.08359.98
LST-AI0.5 ×\times 0.5 ×\times 0.5 mm 315.13134.71
UNETR0.5 ×\times 0.5 ×\times 0.5 mm 349.42309.40
UNETR1.0 ×\times 1.0 ×\times 1.0 mm 322.19238.66
UNETR1.5 ×\times 1.5 ×\times 2.0 mm 348.78302.21
SegFormer0.5 ×\times 0.5 ×\times 0.5 mm 318.48248.73
Table 3: Boundary metrics across models and resolutions: voxel-wise 95th percentile Hausdorff Distance (HD95) and lesion-wise HD95 in the native 0.5 mm 3 7T space.
Figure 1: Comparison of segmentation methods. Shown are 7T FLAIR 0.5 ×\times 0.5 ×\times 0.5 mm 3 images from three participants (A, B, and C) in this study and corresponding segmentations performed and/or trained on the 0.5 ×\times 0.5 ×\times 0.5 mm 3 images. Visual inspection highlights qualitative differences between segmentation outputs from each method. Manual (red) masks indicate the manually annotated image by an expert rater, which acted as the ‘gold standard’ to which quantitative comparisons were performed. LST LPA appears to erroneously highlight bright cortical boundaries, which are often prominent on 7T FLAIR, as lesion, in addition to choroid plexus (A) and septum pellucidum (B). LST AI avoids many of the errors performed by LST LPA, but misses some lesions (left frontal in A) and underfills others (left frontal in C). UNETR and Segformer performed relatively well compared to manual annotation, in addition to capturing small lesions missed by the manual rater (small right subcortical WML in B and C). However, both occasionally erroneously highlighted false positives due to bright cortical boundaries (left frontoparietal in C).
Figure 2: Handling of image artifacts. Shown are 7T FLAIR images (0.5 ×\times 0.5 ×\times 0.5 mm 3 ) from one participant in this study in which a wraparound artifact, likely from ear tissue, occurs in the region of the pons and cerebellum. Manual annotation did not identify this as lesion, but LST LPA, LST AI, and UNETR masked this as a false positive lesion. Segformer did not.
Figure 3: Comparison of models trained at various image resolutions. Shown are 7T FLAIR images from a participant with MS at the original 0.5 ×\times 0.5 ×\times 0.5 mm 3 resolution and downsampled to 1.0 ×\times 1.0 ×\times 1.0 mm 3 and 1.5 ×\times 1.5 ×\times 2.0 mm 3 resolution (first row), along with a manually drawn lesion mask (red). UNETR segmentation shown in yellow and Segformer segmentation shown in purple. The model trained on images and masks in the 0.5 ×\times 0.5 ×\times 0.5 mm 3 resolution space appear closest in segmentation accuracy to the manual, ‘ground-truth’ masks, with a slight reduction in accuracy seen for downsampled 1.0 ×\times 1.0 ×\times 1.0 mm 3 . Downsampling to 1.5 ×\times 1.5 ×\times 2.0 mm 3 clearly demonstrates segmentation inferiority to either isotropic option.

Explore similar work

CardsList

  1. Efficient Transformer-Based Localized Patch Sampling for Choroid Plexus Segmentation in Multiple Sclerosis

    Jun 2, 2026Po-Jui Lu, Alessandro Cagol, Mario Ocampo-Pineda +12Efficient ViTsMagnetic Resonance Imaging

    Translational Imaging in Neurology (ThINk) Basel, Department of Biomedical Engineering, Faculty of Medicine, University Hospital Basel and University of Basel, Basel, Switzerland · Department of Neurology, University Hospital Basel, Basel, Switzerland · Research Center for Clinical Neuroimmunology and Neuroscience Basel (RC2NB), University Hospital Basel and University of Basel, Basel, Switzerland +4

  2. Converting T1-weighted MRI from 3T to 7T quality using deep learning

    Jul 18, 2025Malo Gicquel, Ruoyi Zhao, Anika Wuestefeld +12Magnetic Resonance ImagingMedical Image Synthesis

    Department of Clinical Sciences Malmö, SciLifeLab, Lund University, Lund, Sweden · Univ Rennes, CNRS, Inria, Inserm, IRISA UMR 6074, Empenn ERL U 1228, Rennes, France · Clinical Memory Research Unit, Department of Clinical Sciences Malmö, Lund University, Lund, Sweden +8

  3. Component-Adaptive and Lesion-Level Supervision for Improved Small Structure Segmentation in Brain MRI

    Apr 9, 2026Minh Sao Khue Luu, Evgeniy N. Pavlovskiy, Bair N. TuchinovMedical Image SegmentationRobust Loss Functions

    The Artificial Intelligence Research Center of Novosibirsk State University, Novosibirsk 630090, Russia