Distance guided generative adversarial network for explainable medical image classifications.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional data augmentation relies on intra-domain knowledge, limiting its effectiveness.
• Existing Generative Adversarial Networks (GANs) offer limited inter-domain sample variety.
• Current methods inadequately describe decision boundaries for binary classification tasks.

Purpose of the Study:

• To propose a novel distance-guided GAN (DisGAN) for enhanced data augmentation.
• To improve the description of decision boundaries in binary classification.
• To generate more varied inter-domain and intra-domain samples.

Main Methods:

• Developed DisGAN by combining vertical distance GAN (VerDisGAN) and horizontal distance GAN (HorDisGAN).
• VerDisGAN conditions inter-domain generation on vertical distances.
• HorDisGAN conditions intra-domain generation on horizontal distances, mapping source images to hyperplane for class-specific regions.

Main Results:

• DisGAN consistently outperforms existing GAN-based augmentation methods in explainable binary classification.
• The proposed method demonstrates superior performance in clarifying decision boundaries.
• VerDisGAN effectively produces class-specific regions by mapping images to the hyperplane.

Conclusions:

• DisGAN offers a significant advancement in data augmentation for binary classification.
• The method provides explainable classification by improving decision boundary descriptions.
• DisGAN is applicable to various classification architectures and shows potential for multi-class extensions.