Hybrid generative adversarial network based on frequency and spatial domain for histopathological image synthesis

Area of Science:

• Computational pathology
• Digital image processing
• Deep learning

Background:

• Histopathological slide preparation is complex and costly.
• Existing deep learning methods for histological image generation often overlook frequency domain information.
• There's a need to leverage complementary spatial and frequency domain data.

Purpose of the Study:

• To develop a generative adversarial network (GAN) for high-quality histological image generation.
• To effectively fuse spatial and frequency domain information using a cross-attention mechanism.
• To improve the realism and detail of generated histopathological images.

Main Methods:

• Proposed a GAN with a cross-attention mechanism for spatial-frequency domain feature fusion.
• Incorporated a variable-window mixed attention module for multi-scale feature extraction.
• Utilized spectral filtering to enhance periodic structure extraction and cross-attention for dynamic feature weighting.

Main Results:

• The model achieved efficient spatial-frequency domain fusion, significantly enhancing image generation quality.
• Demonstrated superior performance over eight state-of-the-art models on the Patch Camelyon dataset across five metrics.
• Generated realistic and detailed histological images by preserving key details and reducing redundancy.

Conclusions:

• The proposed method advances automated histopathological image generation.
• The spatial-frequency fusion approach offers significant improvements in image quality.
• This technique holds potential for future clinical applications in digital pathology.