Evaluation of Deep Learning Model for Malignancy Risk in Incidental Pulmonary Nodules

Overview

This study assessed the performance of a deep learning (DL) model trained on lung cancer screening data to estimate malignancy risk in incidental pulmonary nodules detected in routine clinical CT scans. The DL model was compared to the Brock model, a clinically established risk calculator, using a retrospective dataset of nodules sized 5–15 mm. The DL model demonstrated promising applicability in a heterogeneous clinical setting, potentially improving malignancy risk stratification.

Background

Incidental pulmonary nodules are increasingly detected due to widespread use of computed tomography (CT). While most nodules are benign, some represent early-stage lung cancer, where early detection can reduce mortality. Current guidelines recommend follow-up based on nodule size, morphology, and growth, often supplemented by malignancy risk calculators such as the Brock model. Artificial intelligence, particularly deep learning, has shown high accuracy in lung cancer screening populations but its performance in routine clinical care, with more diverse patient populations and imaging protocols, remains uncertain.

Data Highlights

The study included adult patients with incidental solid or part-solid nodules sized 5–15 mm from a Dutch University Medical Centre between 2000 and 2019. Malignant nodules were confirmed by stage 1 lung cancer diagnosis, and benign nodules were selected based on absence of cancer diagnosis and radiology report analysis. CT scans had slice thickness ≤3 mm and matrix size 512×512 or 1024×1024. Nodules were annotated by an experienced radiologist. The DL model was trained on 16,077 nodules (1,249 malignant) from the National Lung Screening Trial dataset.

Key Findings

• The DL model, trained on lung cancer screening data, was evaluated on a heterogeneous clinical routine care dataset of incidental nodules.
• The DL model showed comparable or superior performance to the Brock model in malignancy risk estimation.
• Inclusion criteria ensured nodules were indeterminate and required follow-up, focusing on nodules sized 5–15 mm.
• CT imaging parameters and patient selection reflected real-world clinical diversity, enhancing generalizability.
• Radiologist annotations provided a robust reference standard for malignancy status.
• The study supports the potential for DL models to reduce unnecessary imaging and improve early lung cancer detection outside screening programs.

Clinical Implications

The DL model may assist clinicians in more accurately stratifying malignancy risk of incidental pulmonary nodules detected in routine CT scans, potentially reducing unnecessary follow-up imaging and associated costs and radiation exposure. Integration of such AI tools could complement existing risk calculators like the Brock model and support earlier lung cancer diagnosis in diverse clinical populations.

Conclusion

This evaluation demonstrates that a deep learning model trained on lung cancer screening data can be effectively applied to incidental pulmonary nodules in routine clinical care, offering improved malignancy risk assessment and supporting clinical decision-making.

References

1. Hendrix et al 2021 -- Dataset and methodology for incidental pulmonary nodules
2. National Lung Screening Trial -- Lung cancer screening data source
3. BTS Guidelines -- Pulmonary nodule management