Effective Immunohistochemistry Pathology Microscopy Image Generation Using CycleGAN.

Area of Science:

• Digital pathology
• Computational imaging
• Medical artificial intelligence

Background:

• Immunohistochemistry (IHC) is crucial for diagnosing difficult tumors like neuroendocrine tumors, but generating IHC images is time-consuming and costly.
• Standard hematoxylin-eosin (H&E) stained slides are readily available but lack the specific protein expression data provided by IHC.
• A gap exists in efficiently obtaining IHC-like information from routinely stained pathology slides.

Purpose of the Study:

• To develop an effective method for generating synthetic immunohistochemistry (IHC) pathology microscopic images from hematoxylin-eosin (H&E) stained images.
• To overcome the time and cost barriers associated with traditional IHC image generation.
• To provide a valuable tool for pathologists by enabling virtual IHC staining from H&E slides.

Main Methods:

• Utilizing CycleGAN as the core architecture for unpaired and unannotated image-to-image translation.
• Incorporating multiple instance learning algorithms to enhance the model's performance.
• Applying principles of conditional Generative Adversarial Networks (GANs) to improve synthetic image quality and accuracy.

Main Results:

• Successfully generated synthetic IHC pathology microscopic images from H&E stained images without requiring annotations.
• Demonstrated good performance in image synthesis, creating visually representative IHC images.
• The proposed method offers a viable alternative to traditional IHC staining for certain diagnostic applications.

Conclusions:

• This study presents the first known attempt at generating IHC pathology microscopic images using deep learning.
• The developed method is effective and efficient, offering significant potential benefits for clinical practice.
• The synthetic IHC images can aid pathologists in diagnosing challenging cases, improving patient care and reducing diagnostic costs.