Clinical Scorecard: Evaluation of a Deep Learning Model Trained for Screening to Assess Malignancy in Incidental Pulmonary Nodules

At a Glance

Key Highlights

• Incidental pulmonary nodules are increasingly detected due to widespread CT use; early detection of malignancy is critical to reduce lung cancer mortality.
• Current management relies on imaging features and risk calculators like the Brock model; deep learning models may improve malignancy risk classification.
• The evaluated deep learning model showed performance equal to expert thoracic radiologists and outperformed the Brock model on clinical routine care data.

Guideline-Based Recommendations

Diagnosis

• Use imaging features such as nodule size, morphology, location, and growth for malignancy risk assessment.
• Apply malignancy risk calculators incorporating clinical and demographic factors (e.g., Brock model).
• Confirm nodule characterization with at least two years of follow-up or histological confirmation.

Management

• Follow-up CT imaging is recommended for indeterminate nodules sized 5–15 mm to monitor growth and characteristics.
• Exclude patients with prior cancer diagnoses or benign nodular diseases from standard malignancy risk pathways.
• Use deep learning models as adjunct tools to improve risk stratification and potentially reduce unnecessary imaging.

Monitoring & Follow-up

• Perform follow-up imaging at intervals guided by nodule risk and guideline recommendations.
• Monitor nodules for changes in size or morphology over time to inform management decisions.

Risks

• Radiation exposure from repeated CT imaging.
• Potential overdiagnosis and overtreatment of benign nodules.
• Variability in CT acquisition and patient populations may affect model performance.

Patient & Prescribing Data

Adults with incidental pulmonary nodules sized 5–15 mm detected on routine clinical CT scans

Deep learning malignancy risk estimation models trained on screening data can be effectively applied to routine clinical care populations to aid in risk stratification and management decisions.

Clinical Best Practices

• Select patients carefully excluding those with prior cancers, benign nodular diseases, or nodules with clearly benign imaging features.
• Use CT scans with appropriate technical parameters (slice thickness ≤3 mm, reconstruction matrix 512×512 or 1024×1024) for accurate assessment.
• Incorporate expert radiologist annotation and multidisciplinary review when possible to confirm nodule characterization.
• Consider integrating deep learning malignancy risk models alongside established clinical risk calculators to optimize patient management.

References

• British Thoracic Society Pulmonary Nodule Guidelines
• National Lung Screening Trial (NLST)
• Netherlands Cancer Registry