Generative Adversarial Network for Medical Images (MI-GAN)

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models require large datasets for optimal performance, but acquiring annotated medical images is costly and time-consuming.
• Small datasets in medical imaging lead to poor generalization and overfitting in deep learning algorithms.
• Manual annotation by medical experts is a bottleneck for supervised medical image analysis.

Purpose of the Study:

• To propose a novel Generative Adversarial Network for Medical Imaging (MI-GAN) to generate synthetic medical images and their corresponding segmented masks.
• To address the challenge of limited annotated data in medical imaging for supervised learning tasks.
• To improve the performance of supervised analysis in medical imaging through data augmentation.

Main Methods:

• Development of a new Generative Adversarial Network architecture, termed MI-GAN, specifically designed for medical imaging applications.
• Utilizing MI-GAN for the synthesis of high-fidelity retinal images and their precise segmentation masks.
• Evaluating the generated synthetic data for its utility in supervised medical image analysis.

Main Results:

• MI-GAN successfully generates synthetic medical images with accurate segmented masks.
• The proposed method demonstrates superior performance in generating precise segmented images compared to existing techniques.
• Achieved state-of-the-art performance with a dice coefficient of 0.837 on the STARE dataset and 0.832 on the DRIVE dataset.

Conclusions:

• MI-GAN offers a viable solution for augmenting limited medical imaging datasets.
• The generated synthetic data can effectively support supervised learning tasks in medical image analysis.
• The proposed approach achieves competitive, state-of-the-art results in retinal image synthesis and segmentation.