Single-view 2D CNNs with fully automatic non-nodule categorization for false positive reduction in pulmonary nodule...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Pulmonary nodule detection is crucial but challenged by false positives and diverse non-nodule appearances.
• Existing 3D convolutional neural networks (CNNs) are computationally intensive, hindering deep network construction.

Purpose of the Study:

• To develop an efficient framework for pulmonary nodule detection, focusing on false positive reduction.
• To overcome the limitations of high computational complexity in 3D CNNs for this task.

Main Methods:

• An ensemble of 2D CNNs using single-view 2D patches for improved computational and memory efficiency.
• Automatic categorization of non-nodules using an autoencoder and k-means clustering for enhanced feature learning.
• Training 2D CNNs with categorized non-nodules to improve representative feature extraction.

Main Results:

• The proposed framework achieved state-of-the-art performance with a competition metric score of 0.922.
• Demonstrated superior performance compared to five other frameworks.
• Exhibited low computational demands with 789K parameters and 1024M floating point operations per second.

Conclusions:

• A novel framework using an ensemble of 2D CNNs with automatic non-nodule categorization for pulmonary nodule detection.
• Achieved high accuracy and efficiency by utilizing 2D CNNs and a specialized training scheme.
• Offers a computationally efficient alternative to 3D CNN-based methods for pulmonary nodule detection.