Neurosurgery Resident University of Toronto Toronto, ON, CA
Introduction: Artificial intelligence (AI) via large language models (LLMs) like ChatGPT is increasingly applied in medicine, particularly in oncology. This study examines ChatGPT’s diagnostic and management capabilities with MRI images of meningiomas and glioblastomas. ChatGPT’s performance in MRI interpretation could potentially support neuro-oncologic decision-making as these tools evolve.
Methods: Using publicly available data from the Brain Tumour Segmentation (BraTS) challenge, 60 cases of meningiomas and glioblastomas were randomly selected, with three orthogonal MRI images representing each tumor. Five patient-simulated questions, combined with symptoms such as headache and seizures, were submitted to ChatGPT. Responses were assessed by neuroradiologists and neurosurgeons.
Results: ChatGPT correctly identified T1-weighted contrast-enhanced MRI sequences in 91.7% of cases and localized tumors in 66.7% of cases. However, tumor laterality accuracy was low, falling below 50%. ChatGPT accurately diagnosed meningiomas in 73.3% of cases and gliomas in 83.3%, though it frequently included other intracranial tumors in differential diagnoses. For 96.7% of cases, ChatGPT recommended further MRI scans, and for symptomatic patients, it advised surgical intervention.
Conclusion : While ChatGPT demonstrated notable accuracy in MRI sequence identification and tumor localization, limitations in diagnostic precision, particularly in tumor laterality and differential specificity, highlight the need for cautious application in clinical practice. Ethical and legal implications, alongside concerns over potential over-reliance, underscore the necessity for human oversight. Future improvements could support ChatGPT’s role in neuro-oncology, with a focus on dynamic patient interactions and improved diagnostic capability as a supplementary tool in medical practice.
.jpg)
