Lung nodule false positive reduction using a central attention convolutional neural network on imbalanced data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Computer-aided detection (CAD) systems are crucial for early lung nodule diagnosis.
• Reducing false positives is a key challenge in pulmonary nodule detection due to visual similarities with benign findings.
• Imbalanced learning datasets pose difficulties for accurate nodule identification.

Purpose of the Study:

• To develop a deep learning model for effective false positive reduction in lung nodule detection.
• To address the challenge of imbalanced datasets in pulmonary nodule classification.
• To enhance the accuracy of computer-aided detection systems for lung nodules.

Main Methods:

• A central attention convolutional neural network on imbalanced data (CACNNID) was proposed.
• Techniques including density distribution, data augmentation, noise reduction, and balanced sampling were employed to handle imbalanced data.
• The model was designed to focus on central information and minimize irrelevant edge features for discriminant feature extraction.

Main Results:

• The CACNNID model achieved a mean sensitivity of 92.64% on the LUNA16 dataset.
• The model demonstrated a specificity of 98.71% and an accuracy of 98.69%.
• An area under the curve (AUC) of 95.67% was obtained, indicating strong performance in distinguishing nodules.

Conclusions:

• The proposed CACNNID model shows satisfactory performance in reducing false positives for lung nodule detection.
• The attention mechanism effectively extracts discriminant features, improving classification accuracy.
• The methods used to address data imbalance contribute to a more robust and reliable detection system.