Large language models as developers of quantification software for bone scintigraphy in transthyretin cardiac amyloidosis: a one-shot comparison on a synthetic DICOM phantom

Abstract Purpose The heart-to-contralateral lung (H/CL) ratio is a key semi-quantitative index of bone scintigraphy in transthyretin cardiac amyloidosis (ATTR-CA), but its thresholds were validated mainly for 99m Tc-pyrophosphate. Testing them with diphosphonates requires a fast, reproducible measurement tool. We assessed whether large language models (LLMs) can develop such a tool in one shot, and verified its measurements against a known ground truth. Methods An identical prompt requesting a measurement-only offline web application (native DICOM parser, constrained circular regions of interest [ROIs], H/CL at 1 h and 3 h, retention index [RI], operator-entered Perugini grade, results export, no diagnostic interpretation) was submitted to Claude Opus 5.5, GPT-5.6 Sol, Gemini 3.1 Pro, DeepSeek-V4.1-Flash and Qwen3.8-Omni-Flash. Tools were tested on a synthetic DICOM phantom of planar thoracic images with four cardiac uptake levels and known ground truth, and scored by two nuclear medicine physicians across seven domains. Results Composite scores (0–100) were 97.5 for Claude Opus 5.5 and GPT-5.6 Sol, 91.8 for Gemini 3.1 Pro, 74.4 for Qwen3.8-Omni-Flash and 22.4 for DeepSeek-V4.1-Flash, whose tool could not read DICOM files. The three best tools reproduced total counts exactly and RI within tolerance in all measurements; their H/CL deviations were explained by operator-chosen ROI size. The Qwen3.8-Omni-Flash tool returned plausible but incorrect ratios. Conclusion In this one-shot comparison, three paid-tier LLMs produced research-grade quantification tools, whereas both free-tier outputs were unreliable. AI-generated nuclear medicine software should be verified against a known ground truth before use, even for research.