Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

TransferX-enabled scan-to-prediction pipeline for pediatric low-grade glioma BRAF mutational subtyping

Link

https://github.com/AIM-KannLab/BRAF_Classification

Indexing

Keywords: Pediatrics, MRI, CNS, Brain/Brain Stem, Oncology, Feature Detection, Diagnosis, Supervised Learning, Transfer Learning, Convolutional Neural Network (CNN), BRAF, BRAF V600E, BRAF fusion, pediatric low-grade glioma, self-supervised learning, Grad-CAM, COMDist, nnU-Net, ResNet-50

Content: MR, NR, OI, PD

RadLex: RID4026, RID7480, RID10796

Author(s)

Divyanshu Tak

Zezhong Ye

Anna Zapaischykova

Yining Zha

Aidan Boyd

Sridhar Vajapeyam

Rishi Chopra

Hasaan Hayat

Sanjay P. Prabhu

Kevin X. Liu

Hesham Elhalawani

Ali Nabavizadeh

Ariana Familiar

Adam C. Resnick

Sabine Mueller

Hugo J. W. L. Aerts

Pratiti Bandopadhayay

Keith L. Ligon

Daphne A. Haas-Kogan

Tina Y. Poussaint

Benjamin H. Kann

Organization(s)

Artificial Intelligence in Medicine Program, Mass General Brigham, Harvard Medical School

Department of Radiation Oncology, Dana-Farber Cancer Institute, Brigham and Women’s Hospital, Boston Children’s Hospital, Harvard Medical School

Department of Radiology, Boston Children’s Hospital, Harvard Medical School

Center for Data-Driven Discovery in Biomedicine, Children’s Hospital of Philadelphia

Department of Neurosurgery, Children’s Hospital of Philadelphia

Department of Radiology, Perelman School of Medicine, University of Pennsylvania

Departments of Neurology, Pediatrics, and Neurologic Surgery, University of California San Francisco

Department of Radiology, Brigham and Women’s Hospital, Dana-Farber Cancer Institute, Harvard Medical School

Department of Radiology and Nuclear Medicine, CalifRIM & GROW, Maastricht University

Department of Pediatric Oncology, Dana-Farber Cancer Institute, Boston Children’s Hospital, Harvard Medical School

Department of Pathology, Dana-Farber Cancer Institute, Boston Children’s Hospital, Harvard Medical School

Version

1.0

Funding

National Institutes of Health (U24CA194354, U01CA190234, U01CA209414, R35CA22052, K08DE030216), National Cancer Institute SPORE (2P50CA165962), European Union – European Research Council (866504), RSNA (RSCH2017), Pediatric Low-Grade Astrocytoma Program at the Pediatric Brain Tumor Foundation, William M. Wood Foundation, Botha-Chan Low Grade Glioma Consortium.

Ethical review

IRB-approved retrospective study at Dana-Farber/Boston Children’s Hospital and Children’s Brain Tumor Network with waiver of informed consent due to use of public datasets and retrospective design.

Date

Published: 2024-03-06

References

[1] Tak D, Ye Z, Zapaischykova A, Zha Y, Boyd A, Vajapeyam S, Chopra R, Hayat H, Prabhu SP, Liu KX, Elhalawani H, Nabavizadeh A, Familiar A, Resnick AC, Mueller S, Aerts HJWL, Bandopadhayay P, Ligon KL, Haas-Kogan DA, Poussaint TY, Kann BH. "Noninvasive Molecular Subtyping of Pediatric Low-Grade Glioma with Self-Supervised Transfer Learning". Radiology: Artificial Intelligence. 2024 May;6(3):e230333.. 2024-03-06. doi:10.1148/ryai.230333. PMID: 38446044. PMCID: PMC11140508.

Model

Architecture

Two-stage pipeline: (1) nnU-Net–based 3D tumor autosegmentation and preprocessing from T2-weighted MRI; (2) three binary 2D ResNet-50–based classifiers with modified fully connected layers (1024-neuron layer + final single-neuron output) trained with TransferX (in-domain transfer learning + self-supervised interclass cross-training), followed by consensus decision logic for final multiclass prediction.

Availability

Open-source code and trained models: https://github.com/AIM-KannLab/BRAF_Classification

Clinical benefit

Noninvasive prediction of BRAF mutational subtype (wild type, fusion, V600E) in pediatric low-grade glioma to assist treatment selection (including BRAF pathway–directed therapies), risk stratification, and clinical trial enrollment when tissue diagnosis is infeasible.

Clinical workflow phase

Clinical decision support systems; potential aid when biopsy/genomic testing is infeasible or in low-resource settings.

Decision threshold

Patient-level probabilities are obtained by averaging section-wise probabilities; classification threshold selected by optimizing the Youden index on the internal test set.

Degree of automation

Fully automated scan-to-prediction pipeline including autosegmentation and classification with consensus logic; outputs prediction and probability without manual segmentation.

Indications for use

Imaging-based classification of BRAF mutational status (wild type, fusion, V600E) in patients (ages ~1–25 years) with pediatric low-grade glioma using pretreatment T2-weighted brain MRI in neuro-oncology settings.

Input

Raw pretreatment T2-weighted brain MRI scans; Stage 1 autosegmentation produces preprocessed image and tumor mask which feed section-wise classifiers.

Instructions

Provide raw T2-weighted brain MRI. The pipeline performs preprocessing/registration/skull-stripping and nnU-Net 3D tumor autosegmentation, extracts axial tumor sections, applies three binary ResNet-50 classifiers trained with TransferX, and aggregates outputs via consensus decision logic to yield final subtype prediction and probability at the patient level.

Limitations

Retrospective, single-development-institution study; limited data; mild performance degradation on external testing likely due to inter-institutional MRI parameter heterogeneity; trained and evaluated only on T2-weighted MRI; potential biases from cohort selection; not prospectively validated; 2D approach with section averaging chosen to mitigate overfitting; model performance may depend on similarity of MRI parameters to training data.

Output

CDEs: RDE748

Description: Patient-level classification of BRAF mutational subtype (wild type, fusion, or V600E) with associated probability; also intermediate binary classifier outputs combined via consensus logic.

Recommendation

Use as an assistive tool for research and clinical decision support where biopsy/genomic testing is infeasible; requires prospective validation before routine clinical deployment.

Regulatory information

Authorization status: Not FDA-cleared; research-use only as presented in the publication.

Reproducibility

External testing performed on CBTN dataset; code, trained models, and statistical analysis scripts publicly available enabling replication; model calibration reported.

Use

Intended: Diagnosis, Procedure selection

Out-of-scope: Other

Excluded: Diagnosis, Detection and diagnosis

User

Intended: Physician, Subspecialist diagnostic radiologist, Researcher

Out-of-scope: Layperson

Excluded: Layperson