External validation of a convolutional neural network artificial intelligence tool to predict malignancy in pulmonary...

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Accurate risk stratification of pulmonary nodules detected via CT is crucial for patient management.
• Artificial intelligence (AI) presents a promising avenue for enhancing the accuracy of malignancy risk prediction in pulmonary nodules.
• This study compares a novel AI algorithm, the lung cancer prediction convolutional neural network (LCP-CNN), against the established Brock University model.

Purpose of the Study:

• To evaluate and compare the performance of the LCP-CNN algorithm against the Brock University model in estimating the risk of malignancy in pulmonary nodules.
• To assess the potential of AI in improving the accuracy and efficiency of pulmonary nodule risk assessment.
• To determine if the LCP-CNN can reduce the number of benign nodules requiring further investigation or surveillance.

Main Methods:

• A retrospective dataset of 1397 incidentally detected pulmonary nodules (5-15 mm) from three UK hospitals was utilized for validation.
• Ground truth diagnoses were established through histology, nodule resolution, stability, or expert opinion for pulmonary lymph nodes.
• Model discrimination and performance statistics (area under the curve, false negatives, specificity) were compared between the LCP-CNN and the Brock model at predefined score thresholds.

Main Results:

• The LCP-CNN achieved a higher area under the curve (89.6%) compared to the Brock model (86.8%, p≤0.005), indicating superior discrimination.
• The LCP-CNN identified a larger proportion of nodules scoring below the lowest cancer threshold (24.5%) than the Brock model (10.9%).
• The LCP-CNN demonstrated a lower false-negative rate (0.4%) compared to the Brock model (2.5%), with similar specificity.

Conclusions:

• The LCP-CNN exhibits enhanced discrimination capabilities and a greater ability to identify benign pulmonary nodules without compromising cancer detection compared to the Brock model.
• Implementation of the LCP-CNN has the potential to significantly decrease the number of surveillance CT scans needed, leading to substantial resource savings.
• AI-driven risk prediction offers a valuable tool for optimizing the clinical management of pulmonary nodules.