Publicly Available Datasets for Artificial Intelligence in Neurosurgery: A Systematic Review

Abstract

<b>Introduction</b>: The advancement of artificial intelligence (AI) in neurosurgery is dependent on high quality, large, labeled datasets. Labeled neurosurgical datasets are rare, driven by the high expertise required for labeling neurosurgical data. A comprehensive resource overviewing available datasets for AI in neurosurgery is essential to identify areas for potential model building and areas of needed data construction. <b>Methods</b>: We conducted a systematic review according to PRISMA guidelines to identify publicly available neurosurgical datasets suitable for machine learning. A PubMed search on 8 February 2025, yielded 267 articles, of which 86 met inclusion criteria. Each study was reviewed to extract dataset characteristics, model development details, validation status, availability, and citation impact. <b>Results</b>: Among the 86 included studies, 83.7% focused on spine pathology, with tumor (3.5%), vascular (4.7%), and trauma (7.0%) comprising the remaining. The majority of datasets were image-based, particularly X-ray (37.2%), MRI (29.1%), and CT (20.9%). Label types included segmentation (36.0%), diagnosis (26.7%), and detection/localization (20.9%), with only 2.3% including outcome labels. While 97.7% of studies reported training a model, only 22.6% performed external validation, 20.2% shared code, and just 7.1% provided public applications. Accuracy was the most frequently reported performance metric, even for segmentation tasks, where only 60% of studies used the Dice score metric. Studies often lacked task-appropriate evaluation metrics. <b>Conclusions</b>: We conducted a systematic review to capture all publicly accessible datasets that can be applied to build AI models for neurosurgery. Current datasets are heavily skewed towards spine imaging and lack both clinical patient specific and outcomes information. Provided baseline models from these datasets are limited by poor external validation, lack of reproducibility, and reliance on suboptimal evaluation metrics. Future efforts should prioritize developing multi-institutional datasets with outcome labels, validated models, public access, and domain diversity to accelerate the safe and effective integration of AI into neurosurgical care.

Ähnliche Arbeiten