SHARP-GAN: SHARPNESS LOSS REGULARIZED GAN FOR HISTOPATHOLOGY IMAGE SYNTHESIS.

Area of Science:

• Digital Pathology
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Deep learning for histopathology image analysis requires extensive annotated data, which is costly and time-consuming to acquire.
• Current generative models struggle to produce clear contours for overlapping or touching nuclei in synthetic histopathology images.

Purpose of the Study:

• To develop a novel generative adversarial network (GAN) capable of synthesizing realistic histopathology images with improved nucleus contour definition.
• To address the limitations of existing methods in accurately representing complex cellular structures.

Main Methods:

• A sharpness loss regularized GAN was developed for synthetic histopathology image generation.
• The network utilizes a normalized nucleus distance map instead of a binary mask to encode detailed contour information.
• A sharpness loss function was introduced to enhance the contrast of nucleus contour pixels.

Main Results:

• The proposed method successfully generated realistic histopathology images with clear and well-defined nuclei contours.
• Evaluation using four image quality metrics and segmentation tasks on public datasets confirmed the method's effectiveness.
• Quantitative and qualitative analyses demonstrated superior performance compared to existing approaches.

Conclusions:

• The sharpness loss regularized GAN effectively synthesizes realistic histopathology images with enhanced nucleus contour clarity.
• This approach offers a promising solution for reducing the need for large annotated datasets in histopathology image analysis.
• The method has the potential to advance the development of AI-driven tools for digital pathology.