Related Experiment Video
Updated: Sep 10, 2025
08:41
A MRI-Based Toolbox for Neurosurgical Planning in Nonhuman Primates
Published on: July 17, 2020
5.0K
Publicly Available Datasets for Artificial Intelligence in Neurosurgery: A Systematic Review
Bianca Chan1, Brandon Kim1, Ethan Schonfeld1,2
1Stanford Neurosurgical Artificial Intelligence and Machine Learning Laboratory, Stanford School of Medicine, Stanford University, Stanford, CA 94305, USA.
Journal of Clinical Medicine
|August 28, 2025
Summary
High-quality neurosurgery datasets for artificial intelligence (AI) are scarce, particularly those with outcome labels. Current AI models lack external validation and diverse data, hindering safe integration into clinical practice.
Area of Science:
- Neurosurgery
- Artificial Intelligence
- Machine Learning
- Medical Imaging
Background:
- Advancement in artificial intelligence (AI) for neurosurgery relies on extensive, high-quality labeled datasets.
- Creating such datasets is challenging due to the specialized expertise required for neurosurgical data labeling.
- A comprehensive overview of existing datasets is crucial for identifying gaps and guiding future data collection efforts.
Purpose of the Study:
- To systematically review and characterize publicly available neurosurgical datasets suitable for machine learning applications.
- To identify limitations in current datasets and their impact on AI model development and validation in neurosurgery.
- To provide recommendations for future dataset development to enhance AI integration in neurosurgical care.
Main Methods:
- A systematic literature review adhering to PRISMA guidelines was performed.
- PubMed was searched on February 8, 2025, identifying 267 articles, with 86 meeting inclusion criteria.
- Extracted data included dataset characteristics, model development, validation, availability, and citation impact.
Main Results:
- The majority of studies (83.7%) focused on spine pathology, with limited representation of tumor, vascular, or trauma cases.
- Image modalities included X-ray (37.2%), MRI (29.1%), and CT (20.9%).
- Label types were predominantly segmentation (36.0%) and diagnosis (26.7%), with a scarcity of outcome labels (2.3%).
- External validation was reported in only 22.6% of studies, code sharing in 20.2%, and public applications in 7.1%.
- Accuracy was the most common metric, even for segmentation tasks, with suboptimal evaluation metrics frequently used.
Conclusions:
- Publicly accessible neurosurgical datasets are heavily skewed towards spine imaging and lack essential clinical and outcomes data.
- Existing AI models are limited by insufficient external validation, poor reproducibility, and inappropriate evaluation metrics.
- Future efforts must focus on creating diverse, multi-institutional datasets with outcome labels and validated models to ensure safe AI integration in neurosurgery.
