Automatic detect lung node with deep learning in segmentation and imbalance data labeling.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung nodules are early indicators of lung cancer, with larger nodules posing a higher carcinoma risk.
• Deep learning for medical image analysis faces challenges with imbalanced foreground-background labeling.
• Lung nodules (foreground) typically occupy a small percentage of an image, creating data imbalance.

Purpose of the Study:

• To develop an effective lung nodule segmentation method using a U-Net architecture.
• To address the challenge of imbalanced data in medical image segmentation.
• To improve the accuracy of early lung cancer detection through enhanced nodule identification.

Main Methods:

• Utilized a U-Net-based network architecture (2D U-Net) for lung nodule segmentation.
• Implemented a novel preprocessing technique: complementary labeling, where non-nodule areas are labeled and nodule areas are unlabeled.
• Employed dice coefficient loss as the evaluation function, a standard metric for image segmentation tasks.

Main Results:

• The complementary labeling method demonstrated efficiency, particularly with small datasets.
• The proposed preprocessing approach significantly improved lung nodule detection results.
• The study confirmed the effectiveness of complementary labeling in scenarios with limited data quantity.

Conclusions:

• Complementary labeling is a viable and efficient preprocessing strategy for U-Net in medical image segmentation.
• This method shows promise for improving lung nodule detection in data-scarce environments.
• The integration of ROI segmentation models further enhances the overall performance of lung nodule detection systems.