Sample-Adaptive GANs: Linking Global and Local Mappings for Cross-Modality MR Image Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) are widely used for cross-modality medical image synthesis.
• Existing GAN models learn a single global mapping, which is challenging due to limited data and complex medical image characteristics.
• This approach struggles to optimize synthesis for all samples effectively.

Purpose of the Study:

• To propose novel sample-adaptive GAN models for enhanced cross-modality medical image synthesis.
• To address the limitations of global mapping in existing GANs by incorporating sample-specific characteristics.
• To improve the flexibility and performance of GANs in medical image generation tasks.

Main Methods:

• Developed sample-adaptive GANs with two cooperative paths: a baseline path for global mapping and a sample-adaptive path for local feature extraction.
• The sample-adaptive path models relationships between a sample and its neighbors, using target-modality features as auxiliary information.
• Validated the models on three cross-modality MR image synthesis tasks using public datasets.

Main Results:

• The proposed sample-adaptive GANs significantly outperform state-of-the-art methods in cross-modality MR image synthesis.
• Demonstrated improved synthesis performance by flexibly adjusting to individual sample characteristics.
• Showcased the potential of the sample-adaptive strategy to enhance various backbone GAN models.

Conclusions:

• Sample-adaptive GANs offer a significant advancement in cross-modality medical image synthesis.
• The proposed method provides a flexible and effective approach to optimize GAN performance for diverse medical imaging data.
• The sample-adaptive strategy is a valuable addition that can be readily integrated into existing GAN frameworks.