Pulmonary nodule segmentation with CT sample synthesis using adversarial networks.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Pulmonary nodule segmentation is crucial for lung cancer diagnosis.
• Accurate segmentation aids in early detection and treatment planning.

Purpose of the Study:

• To develop a novel framework for segmenting pulmonary nodules using convolutional neural networks (CNNs).
• To improve the accuracy and robustness of nodule segmentation for lung cancer diagnosis.

Main Methods:

• Utilized a conditional generative adversarial network (cGAN) to synthesize realistic CT images and create a balanced dataset.
• Incorporated semantic and attribute scoring labels to enhance synthetic sample realism.
• Developed a 3D CNN model integrating edge and local binary pattern maps for improved feature learning.
• Employed residual units to accelerate training and boost segmentation accuracy.

Main Results:

• Generated synthetic CT images demonstrated high realism, with an average cosine similarity of 0.9534 to real samples.
• Achieved a Dice coefficient of 0.8483, positive predicted value of 0.8895, sensitivity of 0.8511, and accuracy of 0.9904 for nodule segmentation.
• The framework significantly improved segmentation performance compared to existing methods.

Conclusions:

• The proposed 3D CNN framework, leveraging synthetic data and multi-map inputs with residual learning, offers superior pulmonary nodule segmentation.
• The CT image synthesis method provides realistic and diverse samples, enhancing the training of segmentation models.