OMS-CNN: Optimized Multi-Scale CNN for Lung Nodule Detection Based on Faster R-CNN

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiology

Background:

• Lung cancer detection relies on identifying pulmonary nodules in CT scans.
• Early detection is crucial for reducing lung cancer morbidity and mortality.
• Existing methods like Faster R-CNN can be improved for detecting small nodules.

Purpose of the Study:

• To enhance the Faster R-CNN model for improved pulmonary nodule detection.
• To optimize feature map generation using an advanced convolutional neural network.
• To reduce false positives in lung nodule detection.

Main Methods:

• Implemented an optimized multi-scale convolutional neural network (OMS-CNN) for feature extraction.
• Utilized parameter-setting-free harmony search (PSF-HS) for hyperparameter optimization.
• Employed beetle antenna search (BAS) for initializing kernel filters.
• Integrated multiple 3D deep convolutional neural networks (3D DCNN) for false-positive reduction.

Main Results:

• The OMS-CNN effectively extracted nodule features across various sizes.
• Achieved a sensitivity of 94.89% and a CPM score of 0.892 on LUNA16 and PN9 datasets.
• Demonstrated enhanced detection sensitivity and effective management of false positives.

Conclusions:

• The proposed OMS-CNN technique significantly improves pulmonary nodule detection accuracy.
• The integrated approach offers clinical utility for early lung cancer diagnosis.
• This method serves as a valuable reference for advanced nodule detection systems.