134 CNS-Obsidian: A Neurosurgery-Tailored Vision-Language Model Fine-Tuned Exclusively on Scientific Publications
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
INTRODUCTION: Scientific publications represent a valuable repository of domain-specific knowledge that remains underutilized in training vision-language models (VLMs), particularly for specialized medical fields like neurosurgery where high-quality multimodal data is limited and typically not publicly available. METHODS: We used the 34B parameter version of LLaVA-Next based on Nous Hermes 2 Yi-34B, which combines CLIP-derived image features with language embeddings to create our neurosurgery-specific model, CNS-Obsidian. We implemented a three-stage curriculum training: (1) alignment training on PubMed figure-caption pairs with frozen model parameters except projection layers, (2) medical knowledge integration with unfrozen projection and language model components trained on general medicine PubMed-based conversations, and (3) neurosurgical specialization using domain- and task-specific datasets generated with GPT-4o and Claude Sonnet 3.5. Training was conducted on 104 H100 GPUs using PyTorch FSDP for distributed data parallelization, with paper-level data splitting to prevent information leakage. RESULTS: Our best-performing CNS-Obsidian model statistically matched GPT-4o on held-out GPT-generated multiple-choice questions (p = 0.2161) and significantly outperformed both GPT-4o (p = 0.0011) and Claude Sonnet 3.5 (p = 0.0004) on Claude-generated questions. Zero-shot performance on human-generated CNS-SANS questions improved from 39.81% to 46.81%, though remaining inferior to frontier models (GPT-4o = 65.70%, p < 10^-15). including Claude-generated data in training improved performance on Claude-generated MCQs (p = 0.0427) but showed no significant improvement on GPT-generated MCQs or SANS questions. CONCLUSIONS: CNS-Obsidian achieved performance comparable to frontier models on neurosurgical tasks, highlighting the importance of high-quality in-domain data for specialized AI applications. These findings suggest scientific publications offer an effective solution for developing expert-level AI systems in fields where public data is limited.
Fields of Study & Subjects
Multimodal Machine Learning Applications
Artificial Intelligence in Healthcare and Education
Topic Modeling
Publication details
- Journal / source
- Neurosurgery
- Volume
- 72
- Issue
- Supplement_1
- Pages
- 20–21
- Publication date
- 26 March 2026
- Language
- EN