CSE-GAN: A 3D conditional generative adversarial network with concurrent squeeze-and-excitation blocks for lung...

Area of Science:

• Medical image analysis
• Computer vision
• Artificial intelligence in healthcare

Background:

• Lung nodule segmentation is vital for early lung cancer diagnosis and patient survival.
• Convolutional Neural Networks (CNNs) show promise in medical image analysis but face challenges like data scarcity and class imbalance, leading to overfitting.
• Existing CNN-based segmentation methods require improvement to address these limitations effectively.

Purpose of the Study:

• To propose a novel 3D conditional generative adversarial network (3D cGAN) for enhanced lung nodule segmentation.
• To address data scarcity and class imbalance issues that hinder the performance of traditional CNN models.
• To improve the accuracy and reliability of lung nodule segmentation for better early lung cancer detection.

Main Methods:

• Developed a 3D cGAN comprising a U-Net-based generator with a concurrent squeeze & excitation module and a discriminator with spatial and channel excitation.
• Implemented patch-based training to mitigate overfitting issues.
• Evaluated the proposed model on the LUNA16 dataset and a local dataset.

Main Results:

• Achieved a Dice coefficient of 80.74% on the LUNA test set and 76.36% on the local dataset.
• Obtained sensitivities of 85.46% for the LUNA test set and 82.56% for the local dataset.
• Demonstrated significantly improved segmentation performance compared to existing methods.

Conclusions:

• The proposed 3D cGAN approach effectively enhances lung nodule segmentation accuracy.
• The method successfully addresses challenges of data scarcity and class imbalance in medical image segmentation.
• This technique holds potential for improving early lung cancer diagnosis through more precise nodule identification.