DeepFAN, a transformer-based model for human-artificial intelligence collaborative assessment of incidental pulmonary...

The widespread adoption of computed tomography has increased the detection of lung nodules. However, deep learning methods for classification of benign and malignant nodules often fail to comprehensively integrate global and local features, and most of these methods have not been validated through clinical trials. Here we developed DeepFAN, a transformer-based model trained on more than 10,000 pathology-confirmed nodules, and conducted a multireader, multicase clinical trial (Chinese Clinical Trial Registry: ChiCTR2400084624) to evaluate its efficacy in assisting junior radiologists. DeepFAN achieved diagnostic area under the curve (AUC) values of 0.939 (95% CI 0.930-0.948) on an internal test set and 0.954 (95% CI 0.934-0.973) on a clinical trial dataset involving 400 cases across three independent medical institutions. Explainability analysis indicated higher contributions from global than local features. The average performance of 12 readers improved significantly: by 10.9% (95% CI 8.3-13.5%) for AUC, 10.0% (95% CI 8.9-11.1%) for accuracy, 7.6% (95% CI 6.1-9.2%) for sensitivity and 12.6% (95% CI 10.9-14.3%) for specificity (all P < 0.001). Nodule-level interreader diagnostic consistency improved from fair to moderate (overall κ: 0.313 versus 0.421; P = 0.019). These results indicate that DeepFAN can effectively assist junior radiologists and could help to homogenize diagnostic quality and reduce unnecessary follow-up of patients with indeterminate pulmonary nodules.