SuperDiff: A diffusion super-resolution method for digital pathology with comprehensive quality assessment.

Area of Science:

• Digital pathology
• Medical imaging
• Artificial intelligence

Background:

• Whole Slide Images (WSIs) are crucial for disease diagnosis but often suffer from technical limitations.
• Generative Adversarial Networks (GANs) show promise for image super-resolution but face challenges like overfitting in histopathology.
• Existing evaluation metrics are inadequate for assessing histopathology image quality.

Purpose of the Study:

• To introduce SuperDiff, a novel diffusion-based method for generating and evaluating super-resolution histopathology images.
• To address limitations of GANs in histopathology super-resolution and develop histology-specific evaluation metrics.
• To provide a versatile solution for multi-resolution image generation in digital pathology.

Main Methods:

• Developed SuperDiff, incorporating a restoration module for histopathology priors and a controllable diffusion module.
• Curated two histopathology datasets for training and evaluation.
• Proposed a comprehensive evaluation strategy using full-reference and no-reference metrics.

Main Results:

• SuperDiff demonstrated superior performance compared to state-of-the-art GAN-based methods on multiple datasets.
• The method effectively handles multi-resolution generation from varied input sizes.
• The proposed evaluation strategy provides robust assessment of digital pathology image quality.

Conclusions:

• SuperDiff offers a significant advancement in histopathology image super-resolution.
• The method provides valuable support for diagnostic processes in digital pathology.
• SuperDiff's diffusion-based approach overcomes limitations of GANs in this domain.