PULMONARY nodule assessment on CT scans is evolving as artificial intelligence tools demonstrate measurable improvements in diagnostic performance, according to a recent multi-reader, multi-case clinical trial evaluating a transformer-based model known as DeepFAN.
Pulmonary Nodule Diagnosis Improved With AI
The increasing use of CT imaging has led to a surge in incidental pulmonary nodule detection, presenting ongoing challenges in distinguishing benign from malignant findings. Traditional deep learning models have struggled to integrate both global and local imaging features effectively, limiting their clinical reliability.
DeepFAN was developed to address these limitations using a transformer-based architecture trained on more than 10,000 pathology-confirmed nodules. In testing, the model achieved a diagnostic area under the curve of 0.939 on an internal dataset and 0.954 in a clinical trial involving 400 cases across three independent institutions, indicating strong discriminative performance.
Explainability analysis revealed that global imaging features contributed more substantially to model predictions than local features, suggesting a broader contextual understanding of pulmonary nodules may be critical for accurate classification.
AI Enhances Radiologist Performance and Consistency
The clinical trial focused on human-artificial intelligence collaboration, particularly among junior radiologists. When assisted by DeepFAN, reader performance improved significantly across multiple metrics. Average diagnostic accuracy increased by 10.0%, sensitivity by 7.6%, and specificity by 12.6%, with all improvements reaching statistical significance.
Importantly, diagnostic consistency among readers also improved. Inter-reader agreement increased from a fair level to moderate, indicating that the model may help standardize interpretation and reduce variability in pulmonary nodule assessment.
These findings highlight the potential of AI-assisted workflows to enhance both individual and collective diagnostic performance in clinical practice.
Implications for Pulmonary Nodule Management
Beyond improving accuracy, the model may have practical implications for patient management. By increasing confidence in nodule classification, DeepFAN could reduce unnecessary follow-up imaging and invasive procedures associated with indeterminate findings.
The results position transformer-based AI models as a promising tool for supporting radiologists, particularly in high-volume settings where incidental pulmonary nodules are frequently encountered.
Reference
Zhu Z et al. DeepFAN, a transformer-based deep learning model for human-artificial intelligence collaborative assessment of incidental pulmonary nodules in CT scans: a multi-reader, multi-case trial. arXiv. 2026;DOI:10.48550/arXiv.2603.25607.
Featured Image: WesJVR/peopleimages.com on Adobe Stock.
