Fidelity imposed network edit (FINE) for solving ill-posed image reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Deep learning (DL) excels at solving inverse problems in medical imaging, outperforming traditional methods.
• Supervised DL models struggle with data outside their training distribution, like rare pathologies.
• Existing DL reconstructions may not leverage the physical imaging model, limiting potential performance gains.

Purpose of the Study:

• To introduce Fidelity Imposed Network Edit (FINE), a novel DL approach for medical image reconstruction.
• To enhance the robustness and accuracy of DL reconstructions by incorporating the forward physical model.
• To adapt pre-trained DL networks to specific testing datasets without retraining.

Main Methods:

• Developed Fidelity Imposed Network Edit (FINE), a method to modify pre-trained DL network weights.
• Utilized an unsupervised fidelity loss function based on the forward physical model to guide network adaptation.
• Applied FINE to quantitative susceptibility mapping (QSM) and under-sampled MRI reconstruction.

Main Results:

• FINE demonstrated improved reconstruction accuracy in both QSM and under-sampled MRI.
• The method effectively adapts DL models to individual test cases, enhancing generalization.
• Unsupervised adaptation using the physical model proved beneficial for inverse problem solving.

Conclusions:

• FINE offers a robust solution for improving DL-based medical image reconstruction.
• The approach enhances DL model performance by integrating physical constraints and adapting to test data.
• FINE shows promise for advancing neuroimaging techniques like QSM and accelerated MRI acquisition.