Can AI Outperform Human Radiologists in Diagnostic Reasoning?

Enhancing Clinical Workflows with AI

The potential of LLM diagnostic reasoning in modern clinical workflows has sparked significant interest. Specifically, technical and regulatory constraints often limit the direct application of large language models in radiology. However, reader-mediated text-based workflows can provide a highly practical alternative for busy clinicians. Consequently, researchers evaluated how reader-generated text descriptions affect overall artificial intelligence performance. In addition, they measured how radiologic expertise influences the final clinical decisions.

How Expertise Affects LLM Diagnostic Reasoning

In this study, ten readers independently interpreted ninety-three complex thoracic imaging cases. Specifically, the group included five experienced thoracic radiologists and five radiology residents. During the first session, readers selected their primary diagnosis and wrote a free-text description of key findings. Subsequently, the Gemini 3.0 Pro model analyzed these text descriptions without seeing the actual images. Ultimately, the system ranked five potential differential diagnoses based on the provided text. Therefore, the diagnostic accuracy heavily depended on the quality of the initial human description.

Key Findings From the Study

Interestingly, the model achieved fifty-two point seven percent accuracy when analyzing raw images directly. However, the model achieved sixty-three point nine percent accuracy when utilizing reader-generated descriptions. Furthermore, descriptions from experienced radiologists yielded higher accuracy than descriptions from residents. Specifically, radiologist-guided inputs resulted in sixty-seven point three percent accuracy compared to sixty point four percent for residents. Consequently, reader-mediated workflows can significantly boost diagnostic performance in clinical settings. Ultimately, human-in-the-loop systems show great promise for modern medicine.

Frequently Asked Questions

Q1: What is the main benefit of reader-mediated text-based workflows in radiology AI?

This approach bypasses technical and regulatory constraints. Consequently, it allows clinicians to utilize powerful language models without directly uploading medical images.

Q2: How does radiologist expertise impact LLM diagnostic reasoning?

Expert radiologists provide clearer and more detailed text descriptions. Therefore, the language model can generate significantly more accurate differential diagnoses.

References

1. Song J et al. Human-in-the-Loop Large Language Model-Augmented Diagnostic Reasoning in Thoracic Imaging: Impact of Radiologic Expertise. AJR Am J Roentgenol. 2026 May 20. doi: 10.2214/AJR.26.34999. PMID: 42160120.
2. Kwee TC, Kwee RM. Workload of diagnostic radiologists in the foreseeable future based on recent scientific advances: growth expectations and role of artificial intelligence. Insights Imaging. 2021;12(1):88. doi:10.1186/s13244-021-01031-2.