PathSRGAN: Multi-Supervised Super-Resolution for Cytopathological Images Using Generative Adversarial Network.

Area of Science:

• Digital Pathology
• Medical Imaging
• Artificial Intelligence

Background:

• High-resolution digital slides are crucial for cervical cancer cytopathology screening.
• Acquiring these slides requires expensive equipment and significant time.

Purpose of the Study:

• To develop an efficient super-resolution model for cytopathological images.
• To improve the quality of digital slides for better diagnostic accuracy.

Main Methods:

• Proposed PathSRGAN, a GAN-based progressive multi-supervised super-resolution model.
• Designed a two-stage generator (densely-connected U-Net and residual-in-residual DenseBlock) for 4x to 20x super-resolution.
• Incorporated two supervision terms tailored for cytopathological image characteristics.

Main Results:

• PathSRGAN generated high-quality high-resolution cytopathological images.
• Demonstrated superiority over mainstream CNN-based and GAN-based super-resolution methods.
• Reconstructed images significantly improved computer-aided diagnosis accuracy.

Conclusions:

• PathSRGAN effectively addresses the limitations of high-end imaging equipment in cytopathology.
• The model's performance can enhance diagnostic accuracy and potentially increase screening penetration in resource-limited areas.