Predicting Malignancy and Invasiveness of Pulmonary Subsolid Nodules on CT Images Using Deep Learning

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Pulmonary subsolid nodules (SSNs) require accurate characterization for appropriate management.
• Distinguishing between benign, malignant, pre-invasive, and invasive SSNs is crucial for treatment planning.

Purpose of the Study:

• To develop and validate a deep learning model for predicting malignancy and invasiveness of SSNs on CT images.
• To compare the diagnostic performance of the deep learning model against human expert readers.

Main Methods:

• Retrospective collection of 2,614 patient CT scans with SSNs.
• Development of a 3D convolutional neural network (3D CNN) model for malignancy and invasiveness prediction.
• Comparison of the 3D CNN model's performance with two thoracic radiologists using ROC analysis.

Main Results:

• The 3D CNN model achieved an AUC of 0.913 for benign vs. malignant classification, outperforming radiologists (AUC: 0.846).
• The model demonstrated high sensitivity (86.1%) and specificity (83.8%) for malignancy detection.
• For pre-invasive vs. invasive classification of malignant SSNs, the 3D CNN achieved an AUC of 0.908 with 87.4% sensitivity and 80.8% specificity.

Conclusions:

• The deep learning model shows significant potential for accurately assessing SSN malignancy and invasiveness.
• This AI tool can assist surgeons in making informed treatment decisions for SSNs.
• The model's performance suggests a valuable role in improving diagnostic accuracy for pulmonary nodules.