[An unsupervised unimodal registration method based on Wasserstein Gan]

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Accurate medical image registration is crucial for diagnosis and treatment planning.
• Existing deep learning methods often require ground truth data or predefined similarity metrics, limiting their applicability.
• Unsupervised methods offer a promising alternative but often struggle with accuracy.

Purpose of the Study:

• To develop an unsupervised unimodal image registration method using Wasserstein Generative Adversarial Networks (WGANs).
• To eliminate the need for ground truth data and preset similarity metrics in the training process.
• To achieve high registration accuracy comparable to or exceeding supervised methods.

Main Methods:

• A WGAN-based network comprising a generation and a discrimination network was employed.
• The generation network learns deformation fields and predicts registered images.
• Adversarial training between the networks drives parameter updates without ground truth data.

Main Results:

• The proposed WGAN method demonstrated superior performance on LPBA40 (brain), EMPIRE10 (lung), and ACDC (heart) datasets.
• Achieved the highest DICE coefficient (DSC) and normalized correlation coefficient (NCC) compared to Affine, Demons, SyN, and VoxelMorph algorithms.
• Outperformed the state-of-the-art unsupervised algorithm, VoxelMorph.

Conclusions:

• The proposed unsupervised WGAN method offers a robust and accurate solution for unimodal image registration.
• This approach significantly advances unsupervised registration by removing reliance on ground truth and similarity metrics.
• The method shows potential for broad application in medical image analysis where labeled data is scarce.