Background:
Cutaneous neurofibromas (cNFs) are benign dermal tumors affecting nearly all individuals with NF1 and may number from a few lesions to thousands. Objective and scalable methods to quantify tumor burden are critical for clinical studies and therapeutic trials, yet current approaches rely on manual counting within small framed regions of skin. These methods are labor-intensive, lack standardization, and susceptible to sampling bias due to heterogeneous tumor distribution.

Methods:
We developed cNF-Vision, a deep learning–based toolkit for automated detection, segmentation, and counting of cNFs. To support model development, we assembled one of the largest lesion-level cNF imaging cohorts to date, comprising 80,392 manually annotated lesions across 425 facial and back images from 372 NF1 patients. To accelerate dataset creation, we implemented a point-and-click workflow with machine learning–assisted segmentation that reduced labeling time up to fivefold. Model performance was evaluated on held-out test images using count concordance and segmentation accuracy. The trained model was applied across the full cohort to characterize spatial patterns and clinical associations.

Results:
In evaluation, cNF-Vision showed strong agreement with manual lesion counts (R² = 0.84) and accurate segmentation (Dice = 0.80). Automated cohort-wide analysis revealed non-uniform spatial cNF distribution, including enrichment along the paraspinal axis. The platform enabled standardized, full-field lesion quantification from routine images, overcoming limitations of framed counting approaches.

Conclusion:
cNF-Vision provides a practical, scalable, and reproducible framework for automated cNF burden quantification from routine images. By generating lesion-level and spatial metrics, this approach supports natural history studies, clinical association analyses, and objective outcome measures in cNF trials.

Background:
MRI-based neurofibroma assessment is increasingly important for neurofibromatosis type 1 (NF1) surveillance, treatment monitoring, and clinical trials, but manual or semi-automated analyses are labor-intensive and reader-dependent. Artificial intelligence (AI) may accelerate this process, however, published methods remain difficult to compare and translate into radiologic workflow. Through our previous work on AI-based neurofibroma segmentation in whole-body MRI, we encountered recurring challenges limiting reproducibility, comparability, and clinical translation. This work aims to define key challenges in AI-assisted neurofibroma analysis beyond model architecture.

Methods:
We analyzed studies relevant to AI-assisted neurofibroma analysis in MRI of patients with NF1, including 28 papers published between 2002 and 2026 and covering neurofibroma imaging, segmentation, and classification. We synthesized these sources into a generalized AI-assisted neurofibroma analysis pipeline, identified methodological challenges, and mapped each challenge to the pipeline stage at which it arises.

Results:
Eight key challenges were identified: structural deficiencies in datasets; imaging heterogeneity and domain shift; lesion inclusion criteria and morphologic complexity; annotation uncertainty and reader variability; inconsistent task definition and output granularity; evaluation protocol deficiencies; limited interpretability and workflow integration; and incomplete reporting and reproducibility. These challenges arise at different pipeline stages and jointly limit generalizability, clinical interpretability, and fair comparability of AI methods.

Conclusion:
The development and application of AI-assisted neurofibroma analysis in MRI are affected by multiple methodological challenges. The findings of this work provide the basis for developing recommendations for standardized, clinically meaningful AI-assisted neurofibroma analysis.

Background: Radiogenomics is a promising approach for improving molecular diagnostics in neuro-oncology. It provides clinicians with quantitative imaging biomarkers that complement conventional qualitative MRI assessment.This is particularly important for rare genetic conditions such as schwannomatosis, where integrating imaging and molecular data can enhance diagnostic accuracy and support personalized treatment strategies.
Methods: The 53 Patients with NF2-associated schwannomatosis who were genetically confirmed to carry pathogenic NF2 mutations and had bilateral intracranial vestibular schwannomas eligible for stereotactic radiotherapy were included in the study. All patients were evaluated at the N.N. Burdenko Neurosurgery Center (Moscow, Russian Federation) according to the updated clinical and molecular diagnostic criteria for schwannomatosis proposed by S. Plotkin and G. Evans (2022). Pre-treatment magnetic resonance imaging (MRI) data were used for radiomic feature extraction. Radiomic analysis was conducted using R software (version 4.2.2) in the RStudio Server environment (version 2022.07.0+548) on an NVIDIA DGX A100 supercomputer. A comprehensive set of radiomic features was extracted, including first-order statistics, gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), and shape-based features (3718 features per region of interest).
Results: The analysis revealed distinct imaging patterns and statistically significant differences in radiomic features associated with specific genotypes and clinical phenotypes of the disease, including notable differences in patients with mosaic mutations and copy number variations.
Conclusion: These findings suggest the potential of radiogenomic approaches for non-invasive molecular characterization; however, further validation on larger cohorts is required to confirm the robustness and generalizability of the proposed algorithm.