Diagnostic Evaluation of Computed Tomography Images in Patients with Suspected Lung Lesions Using GPT-4o: A Comparison of Temperature Settings and Diagnostic Accuracy

Lung cancer remains one of the leading causes of cancer-related deaths worldwide. While computed tomography (CT) imaging plays a crucial role in the detection of lung lesions, the potential use of artificial intelligence (AI) tools in assisting diagnosis is increasingly being explored. This study evaluates the diagnostic performance of GPT-4o, a multimodal AI model, in detecting lung lesions on non-contrast chest CT images of patients admitted to the emergency department. We retrospectively analyzed CT scans of 112 adult patients (56 with radiologically confirmed lung lesions and 56 healthy controls) admitted to the emergency department in 2024. GPT-4o was prompted to interpret four selected CT slices per patient at three different temperature settings (low, medium, high). Diagnoses were compared to radiologist assessments for accuracy. GPT-4o demonstrated varying diagnostic performance depending on temperature settings. Sensitivity and specificity at medium temperature were 62.5% and 64.29%, respectively. The highest accuracy (75%) was achieved at low temperature, while high temperature resulted in reduced performance (accuracy: 43.75%). Misinterpretation of lesion location was more common at higher temperatures. GPT-4o demonstrates moderate diagnostic performance in evaluating non-contrast chest CT scans for lung lesions. Its accuracy is highest when using low temperature settings, which favour more conservative and consistent image interpretations. However, the tool showed limitations in detecting lesion locations and in overall sensitivity and specificity. GPT-4o shows potential as a supportive tool in the detection of lung lesions on CT scans, particularly at low temperature settings. However, its diagnostic performance is moderate and should not replace expert radiologist evaluation. Future improvements and prospective validation studies are needed before routine clinical use.