The Effects of Prompt Engineering on a Multimodal Large Language Model in Neuroradiology: A Case Study of GPT-4o for Intracranial Hemorrhage Diagnosis from Brain CT

초록

Multimodal large language models (MLLMs) connect vision and language, yet their diagnostic output can vary with prompt design. We examined how prompting influences GPT-4o for zero-shot intracranial hemorrhage (ICH) classification on non-contrast head CT. Using 60 axial slices from the CQ500 dataset (10 each: intraparenchymal [IPH], intraventricular [IVH], subdural [SDH], epidural [EDH], subarachnoid [SAH]; and 10 normal; one patient per case), we compared five prompts: free response, minimal lesion query, forced-choice among ICH subtypes, chain-of-thought, and self-correction. Performance was measured on a 3-class task (intra-axial, extra-axial, normal) and a 6-class task (five ICH subtypes plus normal). The forced-choice prompt yielded the best results: accuracy of 0.67 and macro-F1 of 0.65 (3-class), and accuracy of 0.42 and macro-F1 of 0.39 (6-class). Chain-of-thought and self-correction did not improve macro-F1 over forced-choice; macro precision rose slightly while accuracy, macro-F1, and the normal-class F1 decreased. Gains were concentrated in IPH and IVH; EDH/SDH improved little, and SAH remained below chance level. Furthermore, an analysis of linguistic sensitivity revealed a strong correlation (R2 = 0.94) between prediction consistency and diagnostic performance, suggesting that the model's robustness to prompt phrasing is intrinsically linked to the visual saliency of the pathology. These findings indicate that constrained, task-aligned prompting can raise zero-shot performance of a general MLLM, but errors among visually similar subtypes persist. Safe clinical integration will require domain-specific fine-tuning beyond prompt design. © 2026 KSII.

키워드