From CNNs to GANs for cross-modality medical image estimation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Cross-modality image estimation generates images from one medical imaging modality to another.
• Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) are deep learning approaches utilized for this task.
• Existing research predominantly focuses on Magnetic Resonance Imaging (MRI) with Positron Emission Tomography (PET) or Computed Tomography (CT).

Purpose of the Study:

• To review the application of CNNs and GANs in cross-modality medical image estimation.
• To outline recent neural network architectures and synthesis techniques.
• To discuss the motivations and challenges within this field.

Main Methods:

• Review of recent literature on CNNs and GANs for medical image estimation.
• Analysis of image-to-image synthesis constructs used in CNNs and GANs.
• Comparison of GANs and CNNs using metrics for estimated versus actual images.

Main Results:

• Generative Adversarial Networks (GANs) demonstrate greater utility than Convolutional Neural Networks (CNNs) in cross-modality image estimation.
• Detailed review of neural network architectures and synthesis methods.
• Identified key challenges including data registration, patch usage, and loss functions.

Conclusions:

• GANs offer enhanced capabilities for cross-modality medical image synthesis compared to CNNs.
• The field faces challenges in data pairing, network design, and accurate intensity projection.
• Further research is needed to address these limitations for improved medical image estimation.