Active Appearance Model Induced Generative Adversarial Network for Controlled Data Augmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Large, annotated medical datasets are crucial for deep learning but difficult to obtain.
• Existing data augmentation methods, like conditional generative adversarial networks (C-GAN), often produce unrealistic images due to limited input information (e.g., shape only).

Purpose of the Study:

• To introduce A-GAN, an enhanced conditional generative adversarial network (C-GAN) for medical image data augmentation.
• To improve the realism and utility of augmented medical images by incorporating both shape and intensity information.

Main Methods:

• Developed an Active Cell Appearance Model (ACAM) to capture statistical distributions of shape and intensity.
• Integrated ACAM into a C-GAN framework to guide image generation, creating A-GAN.
• Applied A-GAN for data augmentation in adaptive optics retinal imaging for cell analysis.

Main Results:

• A-GAN generated more realistic images compared to standard C-GAN.
• A-GAN achieved stable results in fewer iterations than C-GAN.
• Cell detection and segmentation accuracy improved when using A-GAN for data augmentation compared to C-GAN.

Conclusions:

• A-GAN effectively enhances medical image data augmentation by incorporating statistical appearance models.
• The proposed method shows significant potential for improving deep learning-based medical image analysis tasks.
• A-GAN offers a more robust and accurate approach to generating synthetic medical image data.