Hallucination filtering in radiology vision-language models using discrete semantic entropy.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Natural Language Processing

Background:

• Black-box vision-language models (VLMs) are increasingly used for radiologic image analysis.
• Hallucinations, or inaccurate information generation, pose a significant challenge to VLM reliability in clinical settings.
• Current methods for detecting VLM hallucinations are limited, especially in black-box scenarios.

Purpose of the Study:

• To evaluate the efficacy of discrete semantic entropy (DSE) in identifying questions that may lead to hallucinations in VLMs.
• To determine if DSE-based filtering can improve the accuracy of VLMs in radiologic image-based visual question answering (VQA).

Main Methods:

• A retrospective study utilized two public datasets (VQA-Med 2019 and a diagnostic radiology dataset).
• GPT-4o and GPT-4.1 models answered questions, with DSE computed from semantic response clusters.
• Accuracy was assessed before and after filtering questions with DSE > 0.3.

Main Results:

• Baseline accuracy for GPT-4o was 51.7% and for GPT-4.1 was 54.8%.
• After DSE filtering (DSE > 0.3), accuracy increased to 76.3% for GPT-4o and 63.8% for GPT-4.1.
• Accuracy gains were statistically significant across datasets, even after Bonferroni correction.

Conclusions:

• DSE reliably detects hallucinations in black-box VLMs by quantifying semantic inconsistency.
• This method significantly enhances diagnostic answer accuracy in radiologic VQA.
• DSE provides a practical filtering strategy to improve the safety and trustworthiness of clinical VLM applications.