Beyond Mutual Information: Generative Adversarial Network for Domain Adaptation Using Information Bottleneck...

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer vision

Background:

• Multicenter medical images face domain shift, hindering deep learning model generalization.
• Generative adversarial networks (GANs) offer image-to-image (I2I) translation but struggle with object preservation.
• Existing GANs lack practicality for domain adaptation due to object loss in I2I translation.

Purpose of the Study:

• To propose a novel GAN, IB-GAN, for preserving image objects during cross-domain I2I adaptation.
• To enhance the generalization capabilities of deep learning models in medical imaging.

Main Methods:

• Integrated an information bottleneck (IB) constraint into a cycle-consistency-based GAN.
• Discarded superfluous domain information while maintaining disentangled content features.
• Evaluated IB-GAN on polyp, optic disc/cup, and whole heart segmentation tasks.

Main Results:

• IB-GAN successfully preserves image objects during cross-domain I2I translation.
• Generated realistic translated images across diverse medical imaging modalities.
• Significantly improved the generalization performance of segmentation networks like U-Net.

Conclusions:

• IB-GAN effectively addresses the object preservation challenge in cross-domain I2I translation.
• The proposed method enhances the robustness and applicability of deep learning models for medical image analysis.
• IB-GAN shows promise for improving segmentation accuracy in multicenter medical imaging studies.