Generative modeling for renal microanatomy.

Area of Science:

• Digital Pathology
• Machine Learning
• Medical Imaging

Background:

• Expert annotation and large datasets are costly in digital pathology.
• Pathologists rely on limited stain styles for diagnosis, especially with limited tissue samples.
• Generative models offer a potential solution to these data limitations.

Purpose of the Study:

• To investigate the usability and reliability of generative models in digital pathology.
• To synthesize high-resolution renal microanatomical structures using GANs.
• To explore GANs for stain transformation in histopathological images.

Main Methods:

• Utilized Deep Convolutional Generative Adversarial Network (DCGAN) and Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) for unsupervised synthesis of renal structures.
• Employed CycleGAN for stain transformation, converting Hematoxylin and eosin (H&E) to Periodic acid-Schiff (PAS) and Jones methenamine silver (JMS) to PAS stains.
• Assessed GANs' ability to achieve formalin-fixed, paraffin-embedded tissue image quality.

Main Results:

• GANs successfully synthesized high-resolution renal microanatomical structures.
• The models achieved realistic image quality comparable to formalin-fixed, paraffin-embedded tissues.
• CycleGAN effectively transformed between different histopathological stains (H&E to PAS, JMS to PAS).

Conclusions:

• GANs show promise for generating labeled data and overcoming staining limitations in digital pathology.
• Incorporating prior knowledge of micro-anatomical details is crucial for enhancing GAN accuracy.
• GANs can simulate various staining effects, aiding diagnosis when specific stains are unavailable.