Fig. 9: Schematic of the study design. | npj Digital Medicine

Fig. 9: Schematic of the study design.

From: Diagnostic and interpretive gains from reasoning over conclusions with a large reasoning model in radiology

Fig. 9: Schematic of the study design.

A total of 990 radiology reports of cancer cases were collected from three Chinese medical institutions and one public English corpus, with the imaging findings extracted and input into the open-source LRM DeepSeek-R1 to generate the reasoning processes (DeepSeek-R1 (Rea.)) and final conclusion (DeepSeek-R1 (Con.)). The performance of DeepSeek-R1 (Rea.) was then compared with DeepSeek-R1 (Con.) and two state-of-the-art non-reasoning LLMs (DeepSeek-V3_0324 and GPT-4.5) by three senior radiologists on diagnostic, qualitative, and workflow-related metrics. Human-in-the-loop evaluations were also conducted beyond model-level comparison.

Back to article page