Classification of lung nodules based on the DCA-Xception network.

Area of Science:

• Medical imaging analysis
• Deep learning in oncology
• Computational pathology

Background:

• Deep learning for lung nodule classification typically requires extensive datasets.
• Acquiring sufficient medical samples for training is challenging due to their rarity.

Purpose of the Study:

• To evaluate a novel DCA-Xception network integrated with advanced data augmentation techniques.
• To enhance the performance of classifying benign versus malignant lung nodules.

Main Methods:

• Utilized Wasserstein Generative Adversarial Network (WGAN) and five augmentation methods to expand limited sample sizes.
• Developed a DCA-Xception network featuring adaptive dual-channel feature extraction and a convolutional attention module.
• Trained and validated the network on 274 lung nodules and tested on 52 nodules.

Main Results:

• The DCA-Xception network achieved an accuracy of 83.46% and an Area Under the Curve (AUC) of 0.929.
• Features extracted by the network demonstrated 85.24% accuracy when used with K-nearest neighbor and random forest classifiers.
• The proposed method effectively addressed challenges of sample imbalance and insufficiency.

Conclusions:

• The DCA-Xception network significantly outperforms traditional and pre-trained classification networks for lung nodule classification.
• The integration of data enhancement and a specialized network architecture offers a promising approach for improving diagnostic accuracy in medical imaging.