A teacherless lightweight classification framework for benign and malignant pulmonary nodules based on GAS.

Area of Science:

• Medical Imaging
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models are crucial for pulmonary nodule classification but often require substantial computational resources.
• Existing methods face challenges with high memory usage, computational cost, and large parameter counts, hindering clinical adoption.

Purpose of the Study:

• To develop a lightweight and computationally efficient deep learning framework for classifying benign and malignant pulmonary nodules.
• To address the limitations of existing models by reducing parameter count and memory footprint.

Main Methods:

• Proposed the Ghost-Attention Separation (GAS) network, integrating attention mechanisms, residual learning, and an improved DWSGhost module.
• Employed a teacher-free knowledge distillation strategy to create a highly efficient classification model.
• Utilized depthwise separable convolutions for enhanced performance and reduced complexity.

Main Results:

• The GAS network model contains only 119,245 parameters and occupies 0.45 MB, demonstrating significant computational efficiency.
• Experiments on LIDC-IDRI, LungX Challenge, and Zhengzhou Ninth People's Hospital datasets confirmed the model's effectiveness.
• The proposed lightweight model achieved competitive classification performance against other lightweight approaches.

Conclusions:

• The developed lightweight pulmonary nodule classification framework offers enhanced discriminative power, adaptability, and generalization ability.
• The GAS network provides a computationally efficient solution for pulmonary nodule classification, suitable for clinical settings.
• The integration of novel architectural components and knowledge distillation yields a highly performant and resource-efficient model.