RootletSeg: Multi-Contrast Spinal Rootlet Segmentation

• RootletSeg is a deep learning model that segments both dorsal and ventral C2–T1 spinal nerve rootlets using multi-contrast MRI scans.
• It employs a self-configuring 3D U-Net in nnUNetv2 with robust preprocessing and mixed-contrast batching to achieve Dice scores around 0.65.
• The method enables direct neurological level determination for applications such as lesion classification, neuromodulation therapy planning, and fMRI analysis.

RootletSeg is a deep learning method for the automatic segmentation of spinal nerve rootlets on MRI. In its multi-contrast form, it was developed to segment C2–T1 dorsal and ventral spinal rootlets on 3D isotropic 3T TSE T2-weighted scans and on 7T MP2RAGE contrasts comprising T1-weighted INV1, INV2, and UNIT1 images. The model was developed on 93 MRI scans from 50 healthy adults, with training performed on 76 scans and testing on 17 scans, and it is intended to enable the determination of spinal levels directly from MRI scans for downstream analyses including lesion classification, neuromodulation therapy, and functional MRI group analysis (Krejci et al., 17 Sep 2025).

1. Definition, scope, and nomenclature

RootletSeg denotes a family of segmentation methods centered on rootlet-like anatomical or biological structures, but in current spinal MRI usage the term refers specifically to automatic spinal nerve rootlet segmentation. The 2025 multi-contrast implementation extends earlier cervical dorsal-rootlet segmentation work by moving from a T2w-only setting to a framework that operates across T2w and MP2RAGE contrasts, and by segmenting ventral as well as dorsal C2–T1 rootlets (Krejci et al., 17 Sep 2025).

A key antecedent is the T2w-only spinal cord nerve-rootlet model implemented in the Spinal Cord Toolbox, which was described as a 3D multi-class U-Net in nnU-Net v2 for C2–C8 dorsal rootlets and reported an overall Dice score of $0.67 \pm 0.16$ across levels and test volumes (Valosek et al., 2024). A subsequent line of work used dorsal rootlet segmentation as the basis for non-linear registration to the PAM50 template, showing improved alignment relative to disc-based methods in spinal cord fMRI analysis (Bédard et al., 30 Apr 2025). The multi-contrast RootletSeg model thus sits within a broader progression from T2w dorsal-rootlet labeling toward cross-contrast, broader-coverage rootlet analysis.

The name is not unique to spinal MRI. Separate papers use “RootletSeg” or closely related naming for plant-root image segmentation in rhizotron or minirhizotron imagery, including U-Net-based soil-root segmentation and transfer-learning pipelines for root detection in RGB images (Smith et al., 2019, Xu et al., 2019). Geometric prompting for SAM has also been discussed in the context of root segmentation in rhizotron imagery (Ball et al., 27 May 2025). This naming overlap can cause confusion; in spinal neuroimaging, RootletSeg refers to spinal nerve rootlet segmentation rather than plant-root phenotyping.

2. Imaging data and preprocessing pipeline

The multi-contrast RootletSeg study was retrospective and included MRI scans from two open-access and one private dataset. The imaging protocols comprised 3D isotropic 3T TSE T2-weighted scans and 7T MP2RAGE scans with T1w-INV1, T1w-INV2, and UNIT1 contrasts. The cohort consisted of 93 MRI scans from 50 healthy adults with mean age $28.70$ years $\pm 6.53$ [SD], including 28 males and 22 females (Krejci et al., 17 Sep 2025).

All contrasts underwent the same nnUNetv2 preprocessing pipeline. Volumes were reoriented to Right-to-Left $\times$ Posterior-to-Anterior $\times$ Inferior-to-Superior (RPI), intensity z-score normalization was computed per volume, and images were resampled to $0.7 \times 0.7 \times 0.7$ mm isotropic resolution. For UNIT1, an additional contrast inversion step was applied when bootstrapping the initial MP2RAGE model to approximate T2w-like appearances; subsequent training used the original intensities (Krejci et al., 17 Sep 2025).

The data split was staged rather than monolithic. Training of the final multi-contrast model used 76 scans from 50 subjects with an 80/20% training/validation split in 5-fold cross-validation, followed by a final “production” run on all 76 scans. Testing was performed on 17 held-out scans, comprising 12 MP2RAGE and 5 T2w scans, never seen during training or validation (Krejci et al., 17 Sep 2025).

This design suggests that contrast heterogeneity was treated as a core modeling constraint rather than as an external robustness test. A plausible implication is that contrast invariance was learned primarily through shared preprocessing and mixed-contrast batch exposure rather than through explicit contrast-specific subnetworks.

3. Network architecture and training protocol

RootletSeg was implemented in nnUNetv2 as a self-configuring 3D U-Net-style architecture with 6 encoding/decoding stages. The input patch size in RPI orientation was $128 \times 96 \times 352$ voxels at $0.7$ mm isotropic after resampling. Each run used 1 input channel, so the model processed a single MRI contrast per run, while multi-contrast capability was learned by training on mixed-contrast batches. Feature maps followed the default nnUNet doubling at each downsampling; instance normalization was applied after each convolution; Leaky ReLU was used in all hidden layers; and the final layer was followed by Sigmoid for binary segmentation of rootlets versus background. The loss function combined soft Dice loss and cross-entropy loss, and optimization used stochastic gradient descent with Nesterov momentum, initial learning rate $0.01$, the default nnUNet “poly” scheme, and batch size 2 patches per GPU. The output was a one-channel binary segmentation map of ventral plus dorsal C2–T1 rootlets in the same spatial dimensions as the input patch (Krejci et al., 17 Sep 2025).

Model development proceeded in four sequential stages. First, a T2w single-contrast model for ventral plus dorsal rootlets was trained on 31 T2w scans and tested on 5 T2w scans. Second, an initial MP2RAGE model was trained on 5 inverted-UNIT1 scans and tested on the remaining 14 UNIT1 scans, with the reference standard manually corrected. Third, an MP2RAGE-only model was trained on 45 scans corresponding to 15 subjects and 3 contrasts, with patch size tuned to $128 \times 96 \times 352$ to capture C2 context. Fourth, the multi-contrast RootletSeg model was trained on 76 scans and evaluated on the 17 held-out scans (Krejci et al., 17 Sep 2025).

Online nnUNetv2 augmentation included random spatial transforms consisting of translations, rotations, and scalings; axial and coronal mirroring; Gaussian noise and Gaussian blur; brightness and contrast adjustments; Gamma augmentation; and low-resolution simulation through downsampling and upsampling (Krejci et al., 17 Sep 2025).

Performance was measured with the Dice similarity coefficient,

$28.70$0

where $28.70$1 is the predicted segmentation and $28.70$2 is the ground-truth mask (Krejci et al., 17 Sep 2025).

A potential misconception is to treat the multi-contrast model as a semantic level-labeling network analogous to the earlier T2w SCT model. In the 2025 formulation, the output is binary rootlets-versus-background rather than an eight-class level map (Krejci et al., 17 Sep 2025). Spinal-level determination is therefore a downstream analysis built from the segmentation, not a direct output class assignment.

4. Quantitative segmentation performance

On the 17-scan test set, the reported mean $28.70$3 SD Dice scores by contrast were $28.70$4 for T1w-INV2, $28.70$5 for UNIT1, $28.70$6 for T2w, and $28.70$7 for T1w-INV1 (Krejci et al., 17 Sep 2025). These values place performance across all four tested contrasts in a narrow range centered near $28.70$8.

On the same T2w test scans, a previous open-source T2w-only model achieved $28.70$9 Dice, whereas RootletSeg achieved $\pm 6.53$0 Dice, with $\pm 6.53$1 by Wilcoxon signed-rank test (Krejci et al., 17 Sep 2025). The reported interpretation was that RootletSeg extends to four contrasts with no loss in T2w performance.

Relative to the earlier SCT dorsal-rootlet framework, the quantitative picture is different in output structure rather than purely in headline Dice values. The T2w-only SCT model was multi-class and level-specific for C2–C8 dorsal rootlets, whereas the multi-contrast RootletSeg model is binary and includes ventral plus dorsal C2–T1 rootlets (Valosek et al., 2024, Krejci et al., 17 Sep 2025). Direct metric comparison therefore needs to be read in light of differing label spaces, coverage, and contrast domains.

The contrast ranking in the held-out evaluation shows the highest mean Dice for T1w-INV2 and the lowest for T1w-INV1 (Krejci et al., 17 Sep 2025). This suggests that within MP2RAGE-derived contrasts, representation quality for rootlets is not uniform. A plausible implication is that cross-contrast generalization depends not only on network design but also on contrast-specific depiction of rootlet-CSF and rootlet-cord boundaries.

5. Spinal–vertebral level correspondence analysis

A major function of RootletSeg is to support direct estimation of spinal levels from rootlet anatomy rather than from vertebral landmarks. In the correspondence analysis, data from 120 healthy participants were used under conditions of neutral neck position, good rootlet visibility, and coverage from the pontomedullary junction (PMJ) to T1. Spinal levels were derived from rootlet overlap with a dilated spinal cord mask, while vertebral levels were derived from intervertebral-disc tips projected to the cord centerline. Distances from the PMJ to each level midpoint were measured along the centerline and normalized by participant height. Statistical analysis used Wilcoxon signed-rank tests for level-by-level shifts, non-parametric Bland–Altman analysis, and root mean square distance (RMSD) (Krejci et al., 17 Sep 2025).

Bias and limits of agreement were defined as

$\pm 6.53$2

where $\pm 6.53$3 and $\pm 6.53$4 are the PMJ-to-midpoint distances for vertebral and spinal levels, respectively, and $\pm 6.53$5 is the standard deviation of the differences $\pm 6.53$6 (Krejci et al., 17 Sep 2025).

The median bias increased from $\pm 6.53$7 mm for the C2–C3 pair to $\pm 6.53$8 mm for the T1–T1 pair. RMSD likewise increased caudally, from $\pm 6.53$9 mm at C2–C3 to $\times$0 mm at T1–T1 (Krejci et al., 17 Sep 2025). The reported interpretation was a progressively larger rostro-caudal shift between vertebral and neurological spinal levels at more caudal segments, consistent with prior anatomical studies.

This correspondence result is central to the scientific significance of RootletSeg. It formalizes the distinction between vertebral and spinal coordinates in a way that is directly measurable on MRI once rootlets are segmented. In practical terms, RootletSeg is not merely a segmentation model; it is an anatomical indexing mechanism for neurological level assignment.

6. Software integration, applications, and relation to template-based analysis

RootletSeg is open-source. Code and analysis scripts, including reference-standard creation and correspondence analysis, are publicly available at the ivadomed model release, and the packaged model can be invoked via the Spinal Cord Toolbox command sct_deepseg rootlets in SCT version 7.0 or higher (Krejci et al., 17 Sep 2025). The open-access input datasets cited for the T2w component were OpenNeuro ds004507 and the spine-generic multi-subject dataset, while the private MP2RAGE data were shared upon reasonable request to the corresponding author (Krejci et al., 17 Sep 2025).

The downstream analyses explicitly identified for RootletSeg are lesion classification by neurological level, personalized neuromodulation therapy planning through rootlet-level targeting for epidural or intraspinal stimulation, and fMRI group-level spinal cord studies using rootlet-based registration to templates such as PAM50 for functional activation mapping (Krejci et al., 17 Sep 2025). These uses connect segmentation to spatial normalization and intervention planning rather than limiting it to volumetric delineation.

The relation to rootlet-based registration is particularly important. A separate registration study used dorsal cervical rootlet segmentation to drive non-linear alignment to the PAM50 spinal cord template and reported superior alignment across individuals compared with disc-based registration, as well as increased Z scores and larger activation clusters in task-based fMRI, with the number of active voxels increasing from 3292 to 7978 (Bédard et al., 30 Apr 2025). RootletSeg therefore functions as an enabling component in a larger template-registration ecosystem.

Two clarifications are frequently necessary. First, RootletSeg does not replace vertebral labeling by asserting that vertebral landmarks are anatomically irrelevant; rather, it quantifies the mismatch between vertebral and neurological levels and provides a rootlet-based alternative when neurological localization is the objective (Krejci et al., 17 Sep 2025). Second, the multi-contrast method should not be conflated with the earlier SCT T2w-only dorsal-rootlet model, even though both are part of the same methodological lineage (Valosek et al., 2024, Krejci et al., 17 Sep 2025).

Overall, RootletSeg designates an open-source, nnUNetv2-based multi-contrast 3D CNN for C2–T1 spinal rootlet segmentation whose principal contribution is to make spinal-level determination directly accessible from MRI scans. Its significance lies not only in mean Dice values near $\times$1 across four contrasts, but also in the shift from vertebral surrogates toward rootlet-derived neurological localization in spinal imaging workflows (Krejci et al., 17 Sep 2025).