Intracranial Hemorrhage Detection and Subtype Classification on CT Imaging Using a Large Language Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate detection of intracranial hemorrhage (ICH) on CT scans is crucial for patient treatment.
• Convolutional neural networks (CNNs) excel at ICH detection, but large language models (LLMs) for direct medical image interpretation are under-explored.

Purpose of the Study:

• To evaluate a general-purpose multimodal LLM's capability in detecting ICH and classifying its subtypes from CT scans.
• To assess the performance of LLMs in direct medical image interpretation tasks.

Main Methods:

• A multimodal LLM was tested on the PhysioNet ICH dataset for binary ICH detection and multi-class subtype classification.
• Axial CT slices were preprocessed, grouped into composite images, and encoded for LLM input.
• Performance was measured using accuracy, precision, recall, F1 score, exact match accuracy, and Hamming score.

Main Results:

• The LLM achieved 0.52 accuracy for binary ICH detection, with low recall (0.14) for positive cases.
• Subtype classification performance was inconsistent, with intraparenchymal hemorrhage having the highest F1 score (0.57).
• Exact match accuracy was 0.06, and the Hamming score was 0.54, indicating limited predictive ability.

Conclusions:

• The LLM demonstrated limited sensitivity for ICH detection and inconsistent subtype classification in a zero-shot setting.
• Current LLMs require domain-specific fine-tuning, larger datasets, and integration with computer vision methods for clinical deployment.
• Further development is necessary to leverage LLMs effectively in medical image analysis.