A Novel Deep Learning-Based (3D U-Net Model) Automated Pulmonary Nodule Detection Tool for CT Imaging

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate detection of pulmonary nodules on computed tomography (CT) is vital for early diagnosis and effective management of lung conditions.
• Current methods rely heavily on radiologist interpretation, which can be time-consuming and prone to variability.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for automated pulmonary nodule detection in CT scans.
• To compare the algorithm's performance against expert radiologist interpretations.

Main Methods:

• Deep learning models were trained using combined public (LUNA) and private datasets from a tertiary cancer center.
• Performance was assessed using sensitivity, false positives per scan, FROC curves, and the CPM score.
• The algorithm's nodule-wise and patient-wise detection capabilities were evaluated.

Main Results:

• The algorithm achieved 90% sensitivity with only 0.3 false positives per scan (CPM score 0.85).
• For patient-level detection, the algorithm demonstrated 95% sensitivity and 100% specificity across 491 test scans.
• These results indicate high accuracy in identifying pulmonary nodules and affected patients.

Conclusions:

• A validated, multi-institutional deep learning algorithm can effectively assist radiologists in pulmonary nodule detection via CT.
• The AI tool can help confirm findings, identify additional abnormalities, and support national lung screening programs.
• This assistive technology has the potential to improve early diagnosis and patient management for lung diseases.