SIAN: STYLE-GUIDED INSTANCE-ADAPTIVE NORMALIZATION FOR MULTI-ORGAN HISTOPATHOLOGY IMAGE SYNTHESIS

Area of Science:

• Digital pathology
• Medical image synthesis
• Computational biology

Background:

• Current deep neural networks struggle with organ-specific histopathology image style generation.
• Accurate delineation of clustered nuclei boundaries remains a challenge in synthesized images.

Purpose of the Study:

• To develop a novel approach for synthesizing realistic histopathology images with organ-specific styles and accurate nuclei boundaries.
• To improve the performance of instance segmentation models by utilizing the generated synthetic data.

Main Methods:

• Proposed a Style-Guided Instance-Adaptive Normalization (SIAN) approach.
• SIAN involves four phases: semantization, stylization, instantiation, and modulation.
• Utilized semantic maps and learned style vectors for image synthesis and integrated geometrical/topological information for nuclei boundary generation.

Main Results:

• Validated on a multiple-organ dataset, demonstrating superior realism compared to four state-of-the-art methods across five organs.
• Generated histopathology images with improved color distributions, textures, and accurate nuclei boundaries.
• Incorporating synthetic images into model training significantly boosted instance segmentation network performance.

Conclusions:

• The SIAN approach effectively synthesizes realistic, organ-specific histopathology images with precise nuclei boundaries.
• The generated synthetic data enhances the training of instance segmentation models, achieving state-of-the-art results.