Interactive-MEN-RT for Meningioma Segmentation

• Interactive-MEN-RT is a specialized interactive segmentation system designed for 3D delineation of meningioma tumors in contrast-enhanced T1-weighted MRI scans.
• It adapts nnInteractive and nnU-Net V2 principles using iterative refinement with point, bounding box, lasso, and scribble prompts to update volumetric segmentations.
• Experimental evaluations show improved Dice and IoU metrics, underscoring its potential to enhance precision in radiotherapy planning.

Interactive-MEN-RT is a domain-specialized interactive medical image segmentation framework developed for clinician-assisted gross tumor volume segmentation of meningiomas for radiotherapy planning. It is positioned as a dedicated Interactive Medical Image Segmentation system for 3D contrast-enhanced T1-weighted MRI, rather than a general-purpose promptable model, and it addresses a safety-critical contouring problem in which under-segmentation risks undertreating tumor while over-segmentation risks unnecessary irradiation of healthy brain tissue and organs at risk. The framework adapts nnInteractive and nnU-Net V2 principles to a 3D meningioma RT setting, using iterative refinement through point annotations, bounding boxes, lasso tools, and scribbles entered on axial slices to update a volumetric segmentation (Lee et al., 1 Oct 2025).

1. Clinical scope and rationale

Interactive-MEN-RT is designed specifically for meningioma gross tumor volume segmentation in radiotherapy workflows. The motivating premise is that meningioma RT contouring is not adequately served by generic interactive segmentation tools because meningiomas are highly heterogeneous in location, shape, and imaging appearance. Lesions may be relatively well circumscribed at the convexity or may involve skull base, falx/parasagittal, or ventricular regions and abut or encase critical neurovascular structures. In that setting, manual contouring is difficult and inconsistent, while fully automatic models remain unreliable in difficult cases (Lee et al., 1 Oct 2025).

The framework therefore adopts a clinician-in-the-loop compromise. It does not replace contour editing with a one-shot automatic prediction, nor does it reduce the problem to 2D slice-wise annotation. Instead, it uses a 3D promptable segmentation model that begins from image data and accepts explicit user corrections through multiple prompt modalities. This specialization is central to the paper’s contribution: the claim is not that Interactive-MEN-RT introduces an entirely new interactive learning paradigm, but that a strong 3D interactive segmentation framework can be materially improved for treatment-grade meningioma delineation through domain-specific adaptation.

A recurrent misconception is to treat the system as a general MRI interactive segmenter. The paper is more specific. The method is trained and evaluated for a single radiotherapy contouring problem, on a dedicated meningioma RT dataset, and with prompt-response behavior tuned to clinician correction patterns likely to arise in meningioma RT planning. Another potential misconception is to read the framework as multi-parametric because the introduction discusses MRI broadly; in the actual experiments it is single-modality, using only 3D contrast-enhanced T1-weighted MRI.

2. Model architecture and inference workflow

Architecturally, Interactive-MEN-RT is built on nnInteractive and nnU-Net V2 principles and uses a 3D U-Net-based design rather than a transformer backbone. The backbone is the nnU-Net Residual Encoder large configuration, referred to as ResEnc-L. Its input representation combines three semantic components: the original 3D MRI volume, previous segmentation results, and interactive guidance signals. The network outputs a binary tumor-versus-background prediction through a sigmoid output layer (Lee et al., 1 Oct 2025).

The preprocessing pipeline follows nnU-Net-style steps. A CE-T1w MRI volume is subjected to Z-score intensity normalization, resampling to 1 mm isotropic spacing, and cropping to reduce excess background while maintaining anatomical context. During training, augmentation includes rotation, scaling, elastic deformation, and intensity perturbations. At inference time, if the clinician is dissatisfied with the contour, additional prompts are entered in the axial view; these prompts are converted into spatial maps, concatenated to the input, and passed through the model again to obtain an updated 3D segmentation.

A key implementation point is that clinician feedback is incorporated directly at inference time as extra spatial conditioning channels rather than through online weight updates. The paper describes the system as iteratively refining segmentations in response to clinician input, but it does not provide a formal recurrent update equation or an optimization-over-prompts procedure. Refinement occurs by repeated forward passes of the promptable network with updated prompt channels and previous segmentation information.

The loss is described at a high level as DiceCELoss, that is, a combined Dice loss and Cross-Entropy loss. The paper does not provide an explicit formula, and it likewise does not provide formal mathematical expressions for prompt encoding, confidence, uncertainty, or iterative correction dynamics. This suggests that the work is primarily an engineering adaptation of an existing interactive segmentation paradigm to a clinically constrained domain, rather than a mathematically novel optimization framework.

3. Interaction design and prompt representation

The interaction design is one of the framework’s defining features. Interactive-MEN-RT supports four prompt modalities: point annotations, bounding boxes, lasso tools, and scribbles. All are entered as intuitive 2D prompts on a standard axial slice, but they guide a 3D segmentation model. The paper states explicitly that each prompt type is encoded into two additional channels, one positive and one negative, corresponding respectively to foreground inclusion and background exclusion. These channels match the spatial dimensions of the input image and are normalized to the range $[0,1]$ (Lee et al., 1 Oct 2025).

Point annotations are intended for localized correction. Positive points identify voxels or regions believed to belong to tumor, and negative points identify background. Bounding boxes define a broad region of interest on a tumor-containing slice and proved especially effective experimentally. Lasso tools are closed-loop contours for precise delineation of irregular tumor boundaries, and the paper presents them as particularly suited for meningiomas with complex shapes. Scribbles are free-form lines for broader corrective strokes than points while remaining quicker than exact contouring.

The prompt simulation strategy is also domain-motivated. Slice selection for all prompt types was weighted by tumor area so that prompts were more likely to occur on slices where the lesion is prominent. Randomized geometry and jitter were used to mimic the diversity of real user behavior. During training-time simulation, point prompts used 1–2 positive points randomly sampled inside the tumor; boxes were generated on a tumor-containing slice with a random margin; lassos were produced as polygons by sampling 4–12 jittered boundary points on a selected slice; and scribbles were created by connecting 2–8 random intratumoral points with randomized order and added jitter or wavy effects.

This prompt model has two implications. First, it is tuned to how clinicians actually correct contours: points for local errors, boxes for coarse localization, lassos for irregular boundaries, and scribbles for broader refinements. Second, its interaction data are simulated from ground truth rather than collected from real clinicians. The model is therefore optimized for plausible prompt behavior, but not yet validated against prospective user interaction logs.

4. Experimental protocol and quantitative performance

The evaluation uses 500 contrast-enhanced T1-weighted MRI scans from the BraTS 2025 Meningioma RT Segmentation Challenge. The authors used only the challenge training set, partitioning it into 400 cases for training and 100 for validation. Extracranial structures were preserved, while facial features were removed by defacing. Evaluation metrics were Dice Similarity Coefficient and Intersection over Union. No surface-distance metrics, Hausdorff distance, calibration measures, or runtime-per-click analyses were reported (Lee et al., 1 Oct 2025).

Baselines included one automatic method and several interactive methods. The automatic baseline was nnUNet without prompts. Interactive baselines were MedSAM2, SAM-Med3D, and nnInteractive. The authors also compared two variants of their own method: training from scratch and transfer learning initialized from pretrained nnInteractive weights. The transfer-learning ablation is one of the more consequential findings because its benefits depended strongly on prompt type.

Without prompts, the automatic nnUNet baseline achieved $65.5 \pm 25.1$ DSC and $53.1 \pm 23.9$ IoU. Under point prompting, Interactive-MEN-RT trained from scratch reached the best Dice at $75.5 \pm 16.1$ and the best IoU among its variants at $62.8 \pm 17.2$, outperforming SAM-Med3D at $74.9 \pm 21.9$ DSC and $60.4 \pm 20.5$ IoU and nnInteractive at $69.7 \pm 22.2$ DSC and $57.2 \pm 22.3$ IoU. Under bounding-box prompting, the transfer-learning version delivered the best overall box result at $77.6 \pm 11.2$ DSC and $65.5 \pm 25.1$0 IoU. Under lasso prompting, the scratch-trained model achieved the best lasso result at $65.5 \pm 25.1$1 DSC and $65.5 \pm 25.1$2 IoU. Under scribble prompting, the scratch-trained model reached $65.5 \pm 25.1$3 DSC and $65.5 \pm 25.1$4 IoU.

The most concise way to read the headline numbers is that the reported “up to $65.5 \pm 25.1$5 Dice and $65.5 \pm 25.1$6 IoU” comes from two different settings: $65.5 \pm 25.1$7 Dice under bounding-box interaction with transfer learning, and $65.5 \pm 25.1$8 IoU under lasso interaction with scratch training.

Prompt modality	Best Interactive-MEN-RT setting	Best reported result
Point	Scratch	$65.5 \pm 25.1$9 DSC, $53.1 \pm 23.9$0 IoU
Bounding box	Transfer learning	$53.1 \pm 23.9$1 DSC, $53.1 \pm 23.9$2 IoU
Lasso	Scratch	$53.1 \pm 23.9$3 DSC, $53.1 \pm 23.9$4 IoU
Scribble	Scratch	$53.1 \pm 23.9$5 DSC, $53.1 \pm 23.9$6 IoU

The transfer-learning ablation is notable because transfer learning from nnInteractive helped for bounding-box prompts, but scratch training often outperformed transfer learning for points, lassos, and scribbles. The authors interpret this as evidence that generic interactive pretraining may bias the model toward certain prompt or anatomical distributions and may not fully adapt to lesion-specific detail in meningioma RT. A plausible implication is that coarse localization can benefit from general prompt priors, whereas boundary-critical meningioma editing profits more from task-specific optimization.

5. Workflow role, clinical usability, and relation to adjacent radiotherapy systems

The system is intended for radiotherapy workflows and is presented as enabling clinicians to delineate target volumes with minimal input, thereby reducing inter-observer variability and improving efficiency. Its support for multiple prompt types is framed as important for real-world adoption because different users and cases favor different interaction styles, and because axial-view interaction aligns with routine radiology and radiotherapy contouring habits. The fact that the framework is volumetric is also significant: it is not merely a 2D interactive editor, but a 3D segmentation tool whose corrections are propagated through a volumetric model (Lee et al., 1 Oct 2025).

In the broader radiotherapy software landscape, Interactive-MEN-RT occupies the delineation layer rather than the treatment-room simulation layer. This distinction matters because “interactive radiotherapy” can also refer to room-level setup rehearsal, collision visualization, and browser-based treatment-room simulation. A different line of work, exemplified by “Interactive X-ray and proton therapy training and simulation,” addresses collision-aware room simulation for X-ray therapy and proton therapy by combining treatment-machine geometry with patient-specific 3D body models, browser deployment, and interactive manipulation of gantry, couch, and collimator states (Hamza-Lup et al., 2018). Interactive-MEN-RT, by contrast, addresses meningioma contour generation on MRI, not room geometry, hardware kinematics, or collision detection.

That distinction also helps clarify what the framework does not attempt to do. It is not a treatment planning system with integrated dosimetry, and it is not a room-level digital twin. Its contribution lies in target-volume delineation. The paper argues that this narrower focus is clinically justified because radiotherapy planning requires more than a plausible coarse mask: it needs clinically precise contours and a practical mechanism for rapid clinician correction.

6. Limitations, unresolved questions, and future directions

The paper is explicit that the evaluation stops short of a real clinician-in-the-loop study. Prompts were simulated from the ground-truth tumor regions on a per-case basis rather than collected from real users. There is no prospective study, no measurement of the number of interactions required per case, no annotation-time comparison against manual contouring, and no direct physician usability trial. There is likewise no runtime analysis in seconds per interaction, no memory benchmark beyond hardware description, and no formal uncertainty-guided prompting experiment (Lee et al., 1 Oct 2025).

Generalization is also unresolved. The model was evaluated on only one dataset from a single benchmark source, and the authors explicitly note concerns about generalizability to other institutions, protocols, or tumor types. The high variance across cases suggests persistent difficulty in hard lesions, which is consistent with the underlying clinical challenge of tumor heterogeneity and complex anatomy. The method also remains limited to external contour delineation of the tumor target on CE-T1w MRI; it does not incorporate multi-parametric imaging in the reported experiments, and it does not provide explicit mechanisms for uncertainty calibration or confidence-aware interaction scheduling.

Implementation details show that the work is practically reproducible but still experimentally bounded. The software stack is PyTorch 1.13. Training used an NVIDIA A6000 GPU with 48 GB memory. The model was trained on 3D patches of size $53.1 \pm 23.9$7 with batch size 8, using SGD with Nesterov momentum, an initial learning rate of $53.1 \pm 23.9$8, and a polynomial decay schedule. Data augmentation included random rotations of $53.1 \pm 23.9$9, scaling from $75.5 \pm 16.1$0 to $75.5 \pm 16.1$1, elastic deformation, and intensity shifts of $75.5 \pm 16.1$2. Public code is available at the GitHub repository named in the paper.

The future directions identified in the paper are correspondingly pragmatic. The authors call for prospective clinical studies and more comprehensive user-experience assessments. A plausible implication is that the next stage of validation should move from simulated prompt-response behavior to measured contouring efficiency, prompt economy, inter-observer variability, and workflow friction under actual radiotherapy contouring practice. Within that trajectory, Interactive-MEN-RT is best understood as a domain-specialized 3D promptable segmentation framework whose main significance lies in demonstrating that disease-specific adaptation can outperform generic interactive baselines in meningioma radiotherapy planning.