Meningioma Radiotherapy Planning (MEN-RT)

• Meningioma Radiotherapy Planning is a specialized process that uses high-resolution MRI and automated deep learning models to delineate tumor and organ-at-risk structures.
• State-of-the-art segmentation pipelines, incorporating volumetric networks and interactive refinement, significantly reduce manual contouring time while improving reproducibility.
• Integration of precise margin expansion and dose optimization techniques enhances treatment quality and minimizes inter-observer variability in therapeutic planning.

Meningioma Radiotherapy Planning (MEN-RT) refers to the specialized process of imaging-based delineation and optimization of target and risk structures for planning external beam radiotherapy or stereotactic radiosurgery for intracranial meningiomas. Recent years have seen major advances in automated and semi-automated segmentation of meningiomas and associated radiotherapy target volumes, driven by large annotated MRI datasets and state-of-the-art deep learning models. MEN-RT incorporates radiographic, radiomic, and clinical protocols to generate reproducible gross tumor volume (GTV) and organs-at-risk (OAR) contours, supports margin expansions for clinical target volume (CTV) and planning target volume (PTV), and underpins subsequent dose planning and optimization workflows.

1. Imaging Datasets and Annotation Protocols

MEN-RT is fundamentally grounded in high-resolution MRI acquired for radiotherapy planning. The largest datasets to date for automated MEN-RT segmentation are the BraTS-MEN-RT 2024 and the BraTS-MEN 2023 collections. The 2024 dataset consists of approximately 700–800 native-space, defaced 3D post-contrast T1-weighted (T1c) MRI studies from six major U.S. academic centers, covering a spectrum of intact and post-operative cases (LaBella et al., 2024). Each case is paired with a rigorously curated single-label "target volume" representing GTV and any at-risk postoperative margins, annotated in full compliance with RTOG 0539/EORTC 22042 protocols. Annotation follows an institutional or consensus-driven sequence: initial labeling (either from clinical plans or nnU-Net-based pre-segmentation), expert manual correction (ITK-SNAP), and radiologist/radiation oncologist signoff. Quality control procedures explicitly reject any cases with incomplete or erroneously anonymized tumors.

MEN-RT datasets used for deep learning segmentation include only the contrast-enhanced T1 MRI to maximize harmonization across sites and facilitate direct clinical translation (Capellán-Martín et al., 16 Dec 2025). Auxiliary multi-sequence (T1, T2, FLAIR) labeling is prevalent in research segmentation (e.g., BraTS 2023) but is not routine for radiotherapy GTV definition (LaBella et al., 2023).

2. Automated Segmentation: Models and Pipelines

The core technical advance in MEN-RT is robust, automated segmentation of meningioma GTVs. Current state-of-the-art architectures for MEN-RT are built on volumetric deep neural networks, including 3D nnU-Net variants, MedNeXt, and Transformer-based encoders (Capellán-Martín et al., 16 Dec 2025, LaBella et al., 2024). For instance, top BraTS challenge entries utilize:

• Preprocessing: DICOM to NIfTI conversion, z-score intensity normalization, patching/cropping to 128³ voxels (native resolution), and removal of small (<25–300 voxels) disconnected clusters during postprocessing.
• Model types: Multi-fold (e.g., 5-fold) ensemble of 3D nnU-Net V2, nnU-Net with residual encoders, and MedNeXt M. Model weights are combined in a lesion-wise, performance-adaptive manner rather than uniform averaging.
• Radiomic-guided subtyping: Prior to training, cases are stratified by unsupervised clustering on radiomic descriptors (shape, first-order, GLCM, GLSZM features) derived from GTV masks. Principal components (explaining 90% of variance) and k-means clustering (optimized by silhouette score) enhance fold balance across meningioma morphological variants (Capellán-Martín et al., 16 Dec 2025).
• Loss function: Combined Dice and cross-entropy or inversely weighted binary cross-entropy (iwBCE), with explicit correction for small/rare lesions (Shirokikh et al., 2019).
• Augmentation: Random flipping, rotation, elastic deformation, scaling, and intensity shifts.
• Training: Patch-wise, batch size 12–16, learning rate schedules with warm restarts, cross-validation, and ensemble inference.

Performance is quantified by the Dice similarity coefficient (DSC), 95th percentile Hausdorff Distance (HD95), lesion-wise sensitivity (recall), and precision. Recent reports show mean ensemble GTV Dice ≈0.80 (validation), ≈0.81 (test), NSD ≈ 0.68 mm, and median validation Dice 0.887 in robust pipelines (Capellán-Martín et al., 16 Dec 2025).

3. Interactive and Clinician-Guided Segmentation

While automated segmentation achieves high average performance, challenging cases (e.g., irregular or skull-base meningiomas, ambiguous margins) benefit from interactive refinement. Interactive-MEN-RT represents a dedicated 3D U-Net-based IMIS tool, permitting four prompt types: points, bounding boxes, lassos, and scribbles, rendered in 2D but processed volumetrically (Lee et al., 1 Oct 2025). The backbone leverages nnU-Net V2 Residual Encoder configuration, accepts the current MRI, previous prediction, and prompt maps as input channels, and is trained with simulated prompt augmentation (randomized point, box, lasso, scribble sampling near tumor borders).

Evaluation on 500 CE-T1w validation cases (BraTS 2025) demonstrated superior performance to both generic and nnInteractive models:

• Dice coefficients: up to 77.6% with bounding box/lasso, 75.5% with points (mean ± SD typically 11–16%).
• IoU: up to 64.8%.
• Prompt efficiency: Improvements of +5–15% Dice over best non-interactive baselines, especially on complex morphologies.

Clinical integration is designed around a rapid workflow: initial auto-segmentation, minimal-click correction, and direct output of DICOM-RT or NIfTI structure sets.

4. Organ-at-Risk Segmentation and Integration

Comprehensive MEN-RT planning requires parallel segmentation of OARs—eyeball, lens, optic nerve, chiasm, pituitary, hippocampus, brainstem, and brain—to enable quantitative dose constraints (Mlynarski et al., 2019). State-of-the-art OAR segmentation leverages 2D U-Nets (simultaneously trained on multiple organs), optimized batch sampling, and a class-adaptive loss with robust handling of missing labels. Anatomical consistency is enforced by:

• Tri-planar ensemble voting and mask fusion.
• Morphological and logical cleanup (e.g., enforcing eyes/lenses relations).
• Graph-based algorithm for optic nerve centerline extraction, anchoring start/end points to eye/chiasm and reconstructing connectivities and tubular shape.

Quantitative evaluation yields mean surface distances for most OARs in the 0.1–0.7 mm range, mean Hausdorff up to 9.8 mm post-cleanup for all structures, and >96% "Acceptable" or "minor edit" ratings by expert radiotherapists (Mlynarski et al., 2019). Manual OAR delineation time is reduced from hours to minutes, and incorporation into MEN-RT pipelines allows direct dose–volume histogram (DVH) calculation, plan constraint enforcement, and sparing of critical structures.

5. Margin Expansions and Target Volume Derivation

Clinical radiotherapy protocols dictate sequential derivation of planning target volumes:

• GTV: For MEN-RT, this is typically the union of all enhancing and non-enhancing tumor (and for pre-operative cases, includes nodular dural tail). In practice, GTV is defined as the union of ET and NEC labels from multi-label segmentations (LaBella et al., 2023).
• CTV: Computed by isotropic margin expansion of GTV by $r_{\mathrm{CTV}}$ (5–10 mm recommended for benign WHO I tumors).

$$\mathrm{CTV} = \{\,x\in\mathbb{R}^3 : \exists\,y\in \mathrm{GTV},\,\|x-y\|\le r_{\mathrm{CTV}}\}$$

• PTV: Further expansion by $r_{\mathrm{PTV}}$ (typically 2–5 mm for setup uncertainty).

$$\mathrm{PTV} = \{\,x : \exists\,y\in \mathrm{CTV},\,\|x-y\|\le r_{\mathrm{PTV}}\}$$

Volume is computed as

$$V(M)=\sum_{i,j,k}M(i,j,k)\;\times\;v_{\mathrm{vox}}$$

where $M$ is the segmentation mask and $v_{\mathrm{vox}}$ the voxel volume.

For SRS, the clinical workflow may use GTV = CTV (no margin), and set PTV = GTV ⊕ 1 mm morphological dilation (Shirokikh et al., 2019).

6. Dose Planning and Optimization

MEN-RT segmentation outputs feed downstream dose planning systems for both external beam radiotherapy and stereotactic radiosurgery. In GK SRS, multi-resolution-level (MRL) inverse planning enables simultaneous optimization of isocenter positions, sector collimator selection, and beam-on durations (Tian et al., 2019). The convex linear programming model minimizes multi-term objectives: target dose coverage/underdose, shell control for selectivity and gradient index, OAR maximums, beam-on time, and isocenter sparsity. The MRL workflow recursively narrows grid spacing (2 mm→1 mm→0.5 mm) and candidate isocenter points, achieving tractable optimization and interactive trade-off tuning. Quantitative results show for meningiomas:

• 100% coverage with improved selectivity (0.73–0.83), reduced gradient index (2.42–2.57 MRL vs. 2.66 manual), substantial brainstem-sparing (up to 7.4 Gy less max dose), and beam-on time reductions (10–20 min shorter). Shot sequencing post-LP groups sector-collimators into deliverable composite shots compatible with GK hardware.

7. Clinical Integration, Benefits, and Limitations

Automated MEN-RT segmentation introduces critical workflow efficiencies:

• Time savings: 40–70% reduction in contouring time; auto-segmentation plus clinician editing reduces cases to <5–6 min (Shirokikh et al., 2019, LaBella et al., 2024).
• Reproducibility: Standardized, objective GTV/OAR segmentation cuts inter-observer variability by 20–40% (LaBella et al., 2023, LaBella et al., 2024).
• Accuracy: State-of-the-art pipelines (ensemble nnU-Net, MedNeXt, IMIS tools) reach GTV Dice 0.80–0.81 (mean), 0.88 (median), boundary error <0.7 mm, and OAR mean surface distance <1 mm (Capellán-Martín et al., 16 Dec 2025, Lee et al., 1 Oct 2025, Mlynarski et al., 2019).
• Downstream impact: Consistent structure sets improve plan quality, permit automated PTV expansions and dose optimization, and support longitudinal volumetric assessment and radiogenomic correlation.

Limitations are recognized:

• Generalizability: Protocol variation and scanner differences introduce "batch effects," partly mitigated by large multi-site data (LaBella et al., 2023).
• Dural tail and thin/invasive extensions: Remain challenging for both automated and expert segmentation—require clinical review before approval (LaBella et al., 2023, Shirokikh et al., 2019).
• Defacing/skull-stripping: Automated pipelines must preserve extracranial extensions and stereotactic references (LaBella et al., 2024).
• Interactive tools: Most published assessments rely on simulated prompts; real-world clinical speedup and usability require further validation (Lee et al., 1 Oct 2025).

Integration best-practices include QA of each exported structure, harmonization with CT simulation for planning, and institutional tuning of margin recipes.

---

References:

(LaBella et al., 2023, LaBella et al., 2024, Capellán-Martín et al., 16 Dec 2025, Lee et al., 1 Oct 2025, Shirokikh et al., 2019, Mlynarski et al., 2019, Tian et al., 2019)