Deformation-Compensated Learning for Image Reconstruction Without Ground Truth.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Traditional deep learning for medical image reconstruction requires high-quality ground-truth images.
• Noise2Noise (N2N) methods offer an alternative by using multiple noisy measurements, but struggle with nonrigid deformations.

Purpose of the Study:

• To develop a method for training deep reconstruction networks using noisy measurements from objects with nonrigid deformations.
• To improve medical image reconstruction quality in the presence of deformations.

Main Methods:

• Proposed Deformation-Compensated Learning (DeCoLearn) to address nonrigid deformations during training.
• Introduced a deep registration module jointly trained with the reconstruction network.
• Trained networks without requiring ground-truth supervision.

Main Results:

• DeCoLearn successfully trained deep reconstruction networks on data with nonrigid deformations.
• Validated the method on simulated and experimental Magnetic Resonance Imaging (MRI) data.
• Demonstrated significant improvements in medical image reconstruction quality.

Conclusions:

• DeCoLearn is an effective approach for training deep reconstruction networks when ground-truth data is unavailable and objects deform.
• The method enhances the applicability of N2N-based techniques to challenging medical imaging scenarios.
• Jointly trained deep registration and reconstruction networks offer a powerful solution for deformation-affected imaging.