Artificial intelligence has enabled powerful new approaches for volumetric assessment and image analysis in neurofibromatosis type 1, offering the potential for faster, more objective, and reproducible quantification of tumor burden and malignant transformation compared to conventional manual methods. However, translating these tools into routine clinical and research practice remains challenging.

Background
Deciphering the molecular landscape aids in patients’ personalized decision-making. Traditional approaches often begin with multi-omics sequencing based on large biological sample cohorts, limiting their application in rare diseases. Using neurofibromatosis type 1 (NF1) as an example, this study aimed to explore an imaging-initiated approach to delineate molecular subtypes in rare disease, where only limited biological samples can be obtained.

Methods
The study cohort retrospectively included 136 NF1 patients in the study cohort who underwent total-body PET/MRI, while 39 new patients (22 males and 17 females) were included in the validation cohort. A K-means unsupervised clustering algorithm was applied to grade the risk of tumor growth and malignant transformation. Morphological characteristics were then analyzed to define the subtypes. Single cell sequencing was conducted to explore the underlying mechanisms.

Results
Imaging parameters have shown strong correlations with tissue and molecular features, notably with PNFs' variety in potential of tumor growth and malignant transformation. Unsupervised K-means clustering identified three PNF clusters with distinct risks of progression supported by histopathological analyses. We then identified eight PNF subtypes, each characterized by unique morphologies and differing progressing potentials. Single-cell sequencing uncovered a population of NESTIN+SOX2+ neural stem cells in high-risk subtype, which were absent in lower-risk PNFs.

Conclusion
Our new classification system may provide a guideline for individualized management of NF1 patients in a non-invasive way. Furthermore, our research introduces an alternative method for deciphering the molecular landscape in rare diseases, starting from accessible and non-invasive imaging features and minimizing the need for large biological sample sizes.

Background: NF1-associated plexiform neurofibromas (PNFs) require lifelong imaging surveillance. Accurate tumor burden assessment and timely identification of lesions suspicious for malignant transformation remain challenging because PNFs are irregular, nerve-associated, and distributed across diverse anatomical regions.

Methods: Two independent MRI-based AI models were developed to support NF1-PNF management. The PNF segmentation model is an anatomy-guided three-dimensional model adapted from SAMI. It uses candidate nerve-path probability maps, local direction vectors, direction-consistency constrained feature propagation, and non-target anatomical response maps to improve lesion continuity while reducing interference from adjacent muscles, vessels, bone, fat, and organs. The benign-malignant differentiation model is a separately developed T2-weighted MRI-based deep-learning model. It incorporates tumor labels and surrounding tissue or organ labels to distinguish benign PNF from malignant peripheral nerve sheath tumor in complex whole-body backgrounds.

Results: The PNF segmentation model is currently under training. Interim results showed that approximately 50% of cases achieved Dice scores above 0.70, supporting the feasibility of anatomy-guided semi-automatic PNF volumetric assessment, while further optimization is ongoing. For the benign-malignant differentiation model, 347 MRI image sets from seven centers were analyzed. The context-aware one-step model achieved 85.71% accuracy for malignant lesion diagnosis in the validation cohort and 84.75% accuracy in the independent test set, with a false-negative rate of 15.25%. Incorporation of surrounding anatomical labels improved diagnostic performance compared with tumor-label-only models.

Conclusion: These two complementary AI models may support NF1-PNF management by enabling semi-automatic volumetric assessment and imaging-based malignant-risk screening. Prospective validation and workflow integration are warranted.

Background: Optic pathway gliomas associated with neurofibromatosis type 1 (NF1-OPG) are located along the anterior visual pathway (AVP), including the optic nerves (ONs), chiasm, and tracts. To enhance clinical applicability, we introduce CAVS-NF1 (Central AI-enabled Volumetric Service for NF1), an AI-powered web tool designed for automated segmentation and volumetric analysis of longitudinal brain MRIs.

Methods: An automated pipeline utilizing the SwinUNETR, a vision transformer-based AI segmentation model was developed and trained on T1-weighted MRIs of 135 children with NF1-OPG from Children's National Hospital (CNH, GE,n=60) and Children's Hospital of Philadelphia (CHOP, Siemens, n=75). Manual (human) segmentation of the AVP was used for training. Post-segmentation, a template-based registration was employed to isolate the ONs and chiasm. For longitudinal analysis, we aligned the temporal volumes to compute relative changes in substructure volumes.

Results: The AVP segmentation method achieved a Dice score of 79.1 ± 8.2 %, outperforming the inter-observer variability in manual segmentations (Dice: 75.6 ± 6.2 %). The average runtime for a single CAVS-NF1 analysis was 98 ± 3 seconds. The longitudinal analysis workflow demonstrated accurate volume change computations for the AVP with consistent alignment and robust segmentation performance.

Conclusion: CAVS-NF1 enables fast, automated, and accurate segmentation of T1-weighted MRI that produces longitudinal volumetric analysis of NF1-OPG. Our longitudinal volumetric analysis permits a standardized assessment of treatment response ideal for comparing individual patient results to those reported in multi-center clinical trials. This AI-driven web tool is accessible via a free, secure, privacy-preserving HIPAA-compliant web interface, requiring no local installation (https://nf1.hope4kids.io.).

Background: MEK inhibition is increasingly being used to treat plexiform neurofibroma (PN) in patients with Neurofibromatosis Type 1 (NF1). As these targeted therapies enter routine clinical practice, practical and reliable methods for assessing tumor response are needed. This study compared labor-intensive manual segmentation currently considered the gold standard in clinical trials, to simpler measurement methods suitable for clinical practice.
Methods: In this single-center retrospective study, NF1-patients receiving MEK inhibition were included if MRI imaging was available at baseline and after one year of therapy. Target lesions were evaluated using manual volumetric segmentation, bidimensional World Health Organization (WHO) measurements and mean tumor diameter (MTD). Lesions were classified as progressive, stable, or regressive using the consensus 20% volumetric change threshold. Agreement, consistency and measurement time were analyzed.
Results: Thirty-six patients (median age 39.0 years; IQR 35.5; nine <18 years) with 59 PNs were included. WHO and MTD measurements required approximately one minute per lesion, compared to 19.1 (SD 25.4) minutes for volumetric segmentation. In nodular PNs (N = 38), WHO measurements demonstrated substantial agreement with volumetric segmentation (weighted Cohen’s κ = 0.68; 95% CI 0.47–0.88) and identified regression from non-regression with a sensitivity of 0.89 (0.65–0.99) and specificity of 0.80 (0.56–0.94). Multinodular (N=13) and diffuse (=8) PNs were underrepresented, with low agreement.
Conclusion: For nodular PNs, WHO measurements approximate volumetric segmentation and may provide a feasible alternative for clinical practice, while automated segmentation remains unavailable in the majority of centers. To assess detection of progression, we are expanding our analysis using earlier MRI scans.

Background: Optic pathway gliomas (OPGs) are common in neurofibromatosis type 1 (NF1) and are a major cause of childhood visual morbidity. Reliable biomarkers to predict visual deterioration and guide treatment timing are still lacking.

Methods: This retrospective single-center study included 76 untreated children with anterior NF1-associated OPGs undergoing longitudinal ophthalmological assessment and serial brain MRI. Best-corrected visual acuity (VA), global and quadrant-specific retinal nerve fiber layer (RNFL) thickness, and 17 MRI-derived shape-based radiomic features obtained after manual segmentation of the optic nerves and chiasm were analyzed.

Results: No significant association was found between VA and radiomic features across the whole cohort. However, eyes with VA deterioration showed greater increases in maximum 3D and maximum 2D slice diameters, suggesting that diameter-based shape features may be more informative than volumetric measures for identifying visual risk. Mean RNFL thickness correlated with radiomic features only in eyes with chiasmatic involvement. When RNFL was categorized as normal, borderline, or pathological, the superior quadrant showed the strongest associations with tumor volume, surface area, and multiple shape-related features. Eyes with pathological superior RNFL displayed larger, more elongated, and more complex tumors. No longitudinal association emerged between VA deterioration and RNFL thinning, suggesting partial dissociation between structural retinal damage and visual function.

Conclusion: MRI shape-based radiomic features, particularly when integrated with quadrant-specific OCT, provide complementary information beyond conventional volumetric assessment. A multimodal approach combining radiomics, OCT, and VA may improve risk stratification and identify NF1-associated OPGs at increased risk of visual deterioration.

Background:
Hydrocephalus is a recognised but sub optimally characterised complication of neurofibromatosis type 1 (NF1). Most evidence derives from surgical or symptomatic cohorts that over-represent severe disease and under describe milder cases. This study applied text-mixing of radiology reports across an entire NF1 cohort to characterise the spectrum of NF1-related hydrocephalus and ventriculomegaly, including cases conventional cohorts miss.

Methods:
This single-centre retrospective study screened all 3611 magnetic resonance imaging (MRI) brain reports from 965 paediatric-onset NF1 patients within the Manchester Highly Specialised Service for Complex NF1. Patients with documented hydrocephalus, ventriculomegaly, or hydrocephalus-directed intervention underwent detailed analysis. Aetiology, obstruction level, ventricular pattern, detection mode, management, and outcome were drawn from radiology reports and the electronic patient record and were summarised descriptively.

Results:
Fifty patients met inclusion criteria. Hydrocephalus was present in 49 (cohort prevalence 5.1%); one patient had isolated ventriculomegaly. Most were obstructive (96%), the cerebral aqueduct was the commonest site, and triventricular dilatation predominated (71%). Causes were heterogenous: tumoral in 47%, non-tumoural in 37%, with tectal plate lesions the leading single cause. Almost a third (31%) were arrested or compensated, and a similar proportion incidental. Endoscopic third ventriculostomy was the commonest first-line intervention, yet 46% of treated patients required more than one procedure.

Conclusion:
NF1-related hydrocephalus is heterogenous in cause, mechanism, and clinical course rather than a single uniform condition. A substantial proportion is asymptomatic arrested hydrocephalus which required no intervention. These findings support the need for detailed clinical, neuroradiological and multidisciplinary discussions to help inform decision making.

Background:
We aim to evaluate the diagnostic performance of 18F-FDG-PET/CT in identifying malignant transformation of neurofibromas in patients with NF1 managed by the NHS England Highly Specialised Service in Manchester.

Methods:
All patients referred for 18FDG-PET/CT between January 2015 and Dec 2023 were identified. Lesions with histological confirmation were included; lesions without histology were considered benign if clinical and imaging follow-up demonstrated stability. Electronic patient records (EPR) and PACS were manually reviewed to record early and delayed SUVmax, along with mediastinal and hepatic reference activity. Up to five lesions per patient were analysed, prioritising those clinically or metabolically most suspicious. Logistic regression models were constructed using SUVmax, lesion-to-mediastinal SUVmax ratio, and lesion-to-hepatic SUVmax ratio. Clinically relevant thresholds were evaluated, and decision curve analysis was undertaken.

Results:
The lesion-to-hepatic SUVmax ratio demonstrated the best performance, with an AUC of 0.84, outperforming SUVmax alone and the mediastinal ratio. In the final model, higher lesion-to-hepatic SUVmax ratio was strongly associated with malignancy (odds ratio 2.88; 95% CI 1.74–5.46; p < 0.001). Thresholds between 1.8 and 2.0 yielded high negative predictive values (≥0.92) with balanced sensitivity and specificity.

Conclusion:
Lesion-to-hepatic SUVmax ratio is a useful tool in the evaluation of NF1-associated neurofibromas for suspected malignant transformation and outperforms SUVmax alone. Its high negative predictive value may assist as a decision-support tool to guide, with other clinical and MRI parameters, the appropriateness of surgical resection.