HEE-SegGAN: A holistically-nested edge enhanced GAN for pulmonary nodule segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate pulmonary nodule segmentation is crucial for lung cancer screening and disease monitoring.
• Challenges include nodule morphological variability and limited annotated datasets.

Purpose of the Study:

• To propose HEE-SegGAN, a holistically-nested edge-enhanced generative adversarial network, for improved pulmonary nodule segmentation.
• To enhance model robustness and edge segmentation accuracy using a GAN framework integrated with HED-U-Net.

Main Methods:

• Constructed pseudo-color CT images by merging three consecutive slices for spatial continuity.
• Utilized HED-U-Net as the generator and a CNN as the discriminator.
• Incorporated inverted residual modules with channel attention for inter-slice information fusion and feature enhancement.
• Employed deep supervision via HED-U-Net side outputs and minibatch discrimination to mitigate mode collapse.
• Optimized the loss function for improved edge-level detail and precision.

Main Results:

• Ablation experiments on the LUNA16 dataset validated the method's effectiveness.
• Achieved more efficient feature extraction compared to traditional 3D methods, preserving spatial information and reducing computational load.
• Demonstrated improved model stability and segmentation performance through feature matching and minibatch discrimination.

Conclusions:

• HEE-SegGAN shows strong potential for accurate pulmonary nodule segmentation in medical imaging.
• The integrated approach enhances segmentation precision, particularly at nodule edges.
• The method offers efficient feature extraction and reduced computational requirements for lung cancer screening applications.