A Novel Method Based on GAN Using a Segmentation Module for Oligodendroglioma Pathological Image Generation.

Area of Science:

• Digital pathology
• Deep learning
• Medical image analysis

Background:

• Pathological data is scarce, hindering deep learning applications.
• Data augmentation is crucial for training deep learning models with limited datasets.
• Generative models offer potential for synthetic medical data creation.

Purpose of the Study:

• To propose a novel generative model for synthesizing brain tumor pathology data.
• To address challenges in training segmentation models without masked labels.
• To enhance the utility of deep learning in digital pathology.

Main Methods:

• Utilized embedding features from a segmentation module within a general generative model for image synthesis.
• Developed a simple solution for training segmentation models lacking masked labels.
• Employed generative adversarial networks (GANs) or similar generative approaches.

Main Results:

• The proposed method demonstrated improved visual output quality for synthetic pathology images.
• A reduction in the confusion rate across over 70 subjects was observed.
• Quantitative metrics showed limited but notable improvements.

Conclusions:

• The novel generative approach shows promise for augmenting brain tumor pathology datasets.
• The method offers a practical solution for segmentation model training with incomplete data.
• Further research can refine quantitative performance for broader digital pathology applications.