Siamese Cooperative Learning for Unsupervised Image Reconstruction From Incomplete Measurements

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Image reconstruction from incomplete measurements is a fundamental challenge in various imaging modalities.
• Supervised deep learning methods require extensive latent image datasets for training, limiting their applicability in data-scarce scenarios.
• Developing unsupervised methods is crucial for extending deep learning to imaging tasks where acquiring ground truth data is difficult.

Purpose of the Study:

• To propose an unsupervised deep learning framework for image reconstruction that does not require latent images.
• To enable deep learning-based image reconstruction in scenarios where acquiring complete measurement data is challenging.
• To demonstrate the effectiveness of the proposed method across diverse imaging applications.

Main Methods:

• Development of a Siamese network architecture with twin sub-networks operating on complementary data spaces (null and range spaces).
• Implementation of a self-supervised loss function incorporating data consistency and mutual consistency terms.
• Training the deep model using only available measurement data, eliminating the need for latent images.

Main Results:

• The unsupervised Siamese network effectively reconstructs images from incomplete measurements.
• The self-supervised loss ensures accurate reconstruction by enforcing consistency across different data representations.
• Experimental validation on four distinct imaging tasks demonstrates superior performance compared to existing unsupervised solutions.

Conclusions:

• The proposed unsupervised deep learning method offers a viable alternative to supervised approaches for image reconstruction.
• This framework significantly expands the applicability of deep learning in imaging fields with limited or no access to latent data.
• The method shows promise for various imaging applications, particularly those where data acquisition is inherently challenging.