What question did this study set out to answer?

This research aims to evaluate how accurately two AI models interpret neurodegenerative disease diagnostics compared to seasoned experts.

April 26, 2026European Journal of Radiology Artificial Intelligence1 citationsOpen Access

Diagnostic Agreement Between ChatGPT and Nuclear Medicine Experts in the Interpretation of Cerebral 18FFDG-PET/CT for Neurodegenerative Disease

ASArndt‐Hendrik SchievelkampUniversity of Cologne DUDominic UftonUniversity of Cologne JHJan HeilingerUniversity of Cologne

Key Points

This research aims to evaluate how accurately two AI models interpret neurodegenerative disease diagnostics compared to seasoned experts.
Analyzed 100 anonymized cases of neurodegenerative diseases through AI and expert interpretation.
AI models used textual descriptions of imaging findings and clinical history provided as input.
In a reproducibility test, 20 cases were reassessed for run-to-run agreement using an ordinal scale.
ChatGPT-4o and ChatGPT-5 had median agreement scores of 1.00 on a five-level scale.
Main diagnoses were correctly identified in 86% of cases (ChatGPT-4o) and 89% (ChatGPT-5).
Reproducibility rates showed 75% exact agreement for ChatGPT-4o and 55% for ChatGPT-5, with varied consistency levels.

Abstract

Background: Artificial intelligence is a valuable tool in medical imaging and diagnostics.This study assesses the agreement between ChatGPT-4o and ChatGPT-5, two large language models, and expert nuclear medicine physicians in interpreting 18 FFDG-PET/CT findings in neurodegenerative diseases.Methods: 100 anonymized cases were analyzed, comparing AI-generated differential diagnoses with expert reports.Models received only textual descriptions of imaging findings and clinical history extracted from the reports, with patient age provided as an additional input variable.In a 20-case reproducibility subset per model, we re-queried the same cases in a new chat session with the conversation history cleared, using the identical prompt, and assessed run-to-run agreement on a five-level ordinal scale (0, 0.25, 0.5, 0.75, 1).Results: Median agreement scores were 1.00 IQR 0.50-1.00with ChatGPT-4o and 1.00 0.75-1.00with ChatGPT-5.The main diagnosis was correctly identified in 86% (ChatGPT-4o) and 89% (ChatGPT-5) of cases, respectively.Both models performed best in cases with well-defined metabolic patterns but were less accurate for complex metabolic patterns associated with broad differentials.In the reproducibility subset, exact run-to-run agreement was 75% for ChatGPT-4o (quadratic-weighted = 0.48) and 55% for ChatGPT-5 ( = 0.65).The higher for ChatGPT-5 reflects greater consistency on the ordinal scale despite fewer exact matches, with most differences representing one-step shifts.Conclusions: While ChatGPT is not approved for clinical diagnostics, it demonstrated substantial diagnostic alignment with expert physicians when interpreting textual case information.Reproducible outputs highlight its potential as a supportive diagnostic tool.

Demander à l'IA

Bookmark

View Full Paper