Single image super-resolution for whole slide image using convolutional neural networks and self-supervised color...

Area of Science:

• Digital Pathology
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• High-cost whole slide scanners and massive datasets limit digital pathology adoption.
• Existing low-cost scanners produce low-resolution images unsuitable for detailed analysis.

Purpose of the Study:

• To develop a deep learning-based solution for reconstructing high-resolution histology images from low-resolution inputs.
• To improve accessibility and reduce storage/acquisition challenges in digital pathology.

Main Methods:

• Implemented a single image super-resolution (SISR) framework using multi-scale fully convolutional networks.
• Incorporated conditional generative adversarial loss to minimize output image blurriness.
• Employed a progressive training strategy with a normally distributed, increasing scaling factor.

Main Results:

• The SISR framework successfully reconstructed high-resolution images with clinical-level quality.
• A self-supervised color normalization method effectively removed staining variations.
• The combined approach demonstrated generalization on unseen patient tissue cohorts.

Conclusions:

• Deep learning-based SISR offers a viable solution to enhance digital pathology accessibility and quality.
• The proposed method reconstructs clinically relevant high-resolution images from low-cost, low-resolution scans.
• Color normalization improves the robustness and generalizability of the SISR framework for diverse datasets.