Automatic detection of subsolid pulmonary nodules in thoracic computed tomography images

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Subsolid pulmonary nodules have a higher malignancy rate than solid nodules, necessitating precise detection.
• Early and accurate identification of these nodules is critical for lung cancer screening.
• Existing detection methods require improvement for subsolid nodule characterization.

Purpose of the Study:

• To develop and evaluate a computer-aided detection (CAD) system for subsolid pulmonary nodules in CT images.
• To introduce and assess the impact of novel context features for improved nodule classification.
• To optimize and validate the CAD system's performance on a large, multi-center dataset.

Main Methods:

• Development of a CAD system incorporating 128 features, including intensity, shape, texture, and novel context features.
• Training and optimization of the CAD system using a large dataset from a lung cancer screening trial.
• Evaluation of the system's performance on an independent test dataset from a different trial site.

Main Results:

• The proposed CAD system achieved 80% sensitivity with an average of 1.0 false positive per scan.
• The inclusion of novel context features significantly enhanced classification performance.
• Retrospective analysis revealed the CAD system identified subsolid nodules missed by the screening database.

Conclusions:

• The developed CAD system demonstrates high accuracy and efficiency in detecting subsolid pulmonary nodules.
• Novel context features are crucial for improving the performance of CAD systems for subsolid nodules.
• This CAD system shows potential for enhancing lung cancer screening by identifying critical nodules.