UFLoss

UFLoss enhances deep learning-based magnetic resonance imaging (MRI) reconstruction fidelity by introducing a patch-based unsupervised feature loss that preserves fine structures and textures without human annotations.

Key Features:

• Patch-based Unsupervised Feature Loss: A patch-level loss integrated into deep learning frameworks to improve perceptual similarity and preserve high-order statistics.
• Unsupervised Learning: Operates without labeled data, enabling training on unlabeled MRI datasets.
• Instance-wise Discrimination: Maps similar instances to analogous low-dimensional feature vectors to enforce instance-level discrimination during training.
• Low-dimensional Feature-space Loss: Adds an additional loss in the low-dimensional feature space to promote realistic image reproduction from under-sampled or corrupted data.
• Quantitative Evaluation Metrics: Performance assessed using normalized root mean squared error (NRMSE), structural similarity index (SSIM), and the UFLoss metric, with comparisons to methods using ℓ2 loss.

Scientific Applications:

• Medical Imaging: Improves MRI reconstruction fidelity relevant to accurate diagnosis and treatment planning.
• Accelerated 2D and 3D Knee MRI Reconstruction (retrospective under-sampling): Demonstrated application in accelerated knee MRI using retrospective under-sampling techniques.

Methodology:

Train deep learning-based reconstruction frameworks with a patch-based unsupervised feature loss and instance-wise discrimination by mapping to low-dimensional feature vectors, and evaluate using retrospective under-sampling of 2D/3D knee MRI with NRMSE, SSIM, and the UFLoss metric.