Overview

Schema Version

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

Name

Automated Classification of Intramedullary Spinal Cord Tumors and Inflammatory Demyelinating Lesions Using Deep Learning

Link

https://dx.doi.org/10.1148/ryai.210292

Indexing

Keywords: Spinal Cord MRI, Astrocytoma, Ependymoma, Multiple Sclerosis, Neuromyelitis Optica Spectrum Disorder, Deep Learning

Content: MR, NR

RadLex: RID39467, RID4033, RID4055, RID35105, RID10796

Author(s)

Zhizheng Zhuo

Jie Zhang

Yunyun Duan

Liying Qu

Chenlu Feng

Xufang Huang

Dan Cheng

Xiaolu Xu

Ting Sun

Zhaohui Li

Xiaopeng Guo

Xiaodong Gong

Yongzhi Wang

Wenqing Jia

Decai Tian

Xinghu Zhang

Fudong Shi

Sven Haller

Frederik Barkhof

Chuyang Ye

Yaou Liu

Organization(s)

Department of Radiology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China

Department of Neurosurgery, Beijing Tiantan Hospital, Capital Medical University, Beijing, China

Center for Neurology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China

BioMind, Beijing, China

Department of Medical Imaging Product, Neusoft Group, Shenyang, China

China National Clinical Research Center for Neurologic Diseases, Beijing, China

Department of Neurology and Tianjin Neurologic Institute, Tianjin Medical University General Hospital, Tianjin, China

Department of Imaging and Medical Informatics, University Hospitals of Geneva and Faculty of Medicine of the University of Geneva, Geneva, Switzerland

UCL Institutes of Neurology and Healthcare Engineering, London, England

Department of Radiology and Nuclear Medicine, Amsterdam University Medical Centers, the Netherlands

School of Information and Electronics, Beijing Institute of Technology, Beijing, China

Version

1.0

Contact

Corresponding author: Yaou Liu, email: gro.httjb@uoayuil (as provided in article)

Funding

Supported by the National Science Foundation of China (nos. 81870958 and 81571631), the Beijing Municipal Natural Science Foundation for Distinguished Young Scholars (no. JQ20035), the Special Fund of the Pediatric Medical Coordinated Development Center of Beijing Hospitals Authority (no. XTYB201831), and the ECTRIMS-MAGNMIS Fellowship from ECTRIMS (Y.L.).

Ethical review

Study performed in accordance with the Declaration of Helsinki; approved by the animal and human ethics committee of the local institution; written informed consent obtained from all patients.

Date

Updated: 2022-08-24

Published: 2022-09-07

Created: 2021-11-22

References

[1] Zhuo Z, Zhang J, Duan Y, et al.. "Automated Classification of Intramedullary Spinal Cord Tumors and Inflammatory Demyelinating Lesions Using Deep Learning". Radiology: Artificial Intelligence. 2022 Nov;4(6):e210292.. 2022-09-07. doi:10.1148/ryai.210292. PMID: 36523644. PMCID: PMC9745442.

Model

Architecture

Two-dimensional MultiResUNet for lesion segmentation and DenseNet121 for image-based classification.

Availability

Code publicly available at https://github.com/Leezhaohui/spinalcord_classification

Clinical benefit

Assists accurate radiologic diagnosis by segmenting intramedullary spinal cord lesions and differentiating tumors (astrocytoma, ependymoma) from inflammatory demyelinating lesions (MS, NMOSD) and classifying their subtypes.

Clinical workflow phase

Clinical decision support systems; aids radiologic interpretation and differential diagnosis.

Degree of automation

Semi-automated pipeline: automated segmentation with manual verification/correction for poorly segmented lesions, followed by automated classification.

Indications for use

Differential diagnosis of intramedullary spinal cord lesions in patients with suspected tumors (astrocytoma, ependymoma) or inflammatory demyelinating diseases (MS, NMOSD) using clinically available sagittal T2-weighted MR images in a radiology setting.

Input

Sagittal non-contrast T2-weighted MR images (all sections) of the spinal cord; slices with lesions and corresponding lesion masks are used for classification. Contrast-enhanced T1-weighted images were optionally evaluated and improved MS vs NMOSD classification in sensitivity analysis.

Instructions

First segment lesions on sagittal T2-weighted images using the MultiResUNet model; perform manual verification and correct poorly segmented lesions; then use T2-weighted slices containing lesions and corresponding masks as input into the DenseNet121 classifier for: (1) tumor vs demyelination, (2) astrocytoma vs ependymoma, and (3) MS vs NMOSD.

Limitations

Only T2-weighted images used for primary development/validation; segmentation performance for demyelinating lesions was lower and often required manual review/correction; no external testing beyond the single-center prospective cohort; performance degradation observed in pediatric and male subgroups for astrocytoma vs ependymoma; additional lesion types (e.g., spinal cord infarction) not included.

Output

CDEs: RDE746.3, RDE746.0, RDE1455, RDE1451, RDE746, RDE1452

Description: Lesion segmentation masks and categorical classifications for (1) tumor vs demyelination, (2) astrocytoma vs ependymoma, and (3) MS vs NMOSD.

Recommendation

For differentiating demyelinating lesions (MS vs NMOSD), consider adding contrast-enhanced T1-weighted images to improve classification accuracy.

Use

Intended: Decision support, Image segmentation, Diagnosis, Detection and diagnosis

User

Intended: Radiologist, Subspecialist diagnostic radiologist