Selective synthetic augmentation with HistoGAN for improved histopathology image classification.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Histopathological analysis is the gold standard for diagnosing precancerous lesions.
• Automated histopathological classification requires extensive expert annotations, which are costly and time-consuming.
• Accurate classification of image patches is crucial for whole-slide image analysis.

Purpose of the Study:

• To develop a conditional generative adversarial network (GAN), HistoGAN, for synthesizing realistic histopathology image patches.
• To create a novel synthetic augmentation framework that selectively incorporates HistoGAN-generated images.
• To improve the performance of automated histopathological classification by addressing annotation limitations.

Main Methods:

• Designed a conditional GAN model (HistoGAN) to generate class-conditioned synthetic histopathology image patches.
• Developed a selective synthetic augmentation framework that uses confidence and feature similarity for quality assurance.
• Evaluated the framework on cervical and lymph node histopathology datasets.

Main Results:

• HistoGAN successfully synthesized realistic histopathology image patches.
• Selective augmentation with HistoGAN-generated images significantly improved classification accuracy.
• Achieved 6.7% higher accuracy on cervical histopathology and 2.8% higher accuracy on metastatic cancer datasets.

Conclusions:

• HistoGAN and selective synthetic augmentation offer a viable solution to reduce reliance on expert annotations.
• This approach enhances the performance of automated histopathological classification systems.
• The framework provides a quality-controlled method for leveraging synthetic data in medical image analysis.