High fidelity deep learning-based MRI reconstruction with instance-wise discriminative feature matching loss

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) based reconstructions often struggle with fidelity of fine structures and textures.
• Improving image quality in accelerated or corrupted data reconstruction remains a challenge.

Purpose of the Study:

• To enhance reconstruction fidelity of fine structures and textures in deep learning-based reconstructions.
• To introduce a novel unsupervised loss function for improved image realism and detail preservation.

Main Methods:

• Proposed a novel patch-based Unsupervised Feature Loss (UFLoss) incorporated into DL reconstruction frameworks.
• UFLoss provides instance-level discrimination by mapping similar instances to low-dimensional feature vectors, trained without human annotation.
• Applied UFLoss to unrolled networks for accelerated 2D and 3D knee MRI reconstruction.

Main Results:

• UFLoss encourages sharper edges and more faithful contrasts compared to traditional and pure loss-based learning methods.
• Enhanced texture details observed in both 2D and 3D knee MR images.
• Reconstructions with UFLoss showed comparable NRMSE, higher SSIM, and significantly lower UFLoss values.

Conclusions:

• UFLoss enables DL-based reconstruction to achieve more detailed textures, finer features, and sharper edges.
• The method results in higher overall image quality in DL-based reconstruction frameworks.
• Code for UFLoss is publicly available for further research and application.