A new image reconstruction approach that combines locally low-rank modeling with deep learning may improve dynamic contrast-enhanced (DCE) breast MR by delivering higher image quality and substantially faster temporal resolution than conventional techniques, according to a study published in Nature Communications.

The method, called Enhanced Locally low-rank Imaging for Tissue contrast Enhancement (ELITE), was developed by researchers from Weill Cornell Medicine, New York University and the Technion-Israel Institute of Technology. It is designed to address a long-standing tradeoff in breast MR between spatial resolution, temporal resolution and scan time.

The investigators evaluated ELITE using data from 54 patients who underwent clinical breast MR, including 25 with malignant lesions and 29 with benign lesions. The framework combines tissue-specific locally low-rank reconstruction with a residual neural network trained to suppress undersampling artifacts, allowing reconstruction of dynamic images with temporal resolution as short as 1 second while preserving anatomical detail.
Compared with conventional reconstruction methods, ELITE improved lesion conspicuity, reduced background noise and produced higher contrast-to-noise ratios. The approach also shortened reconstruction time relative to standard GRASP-based methods by requiring fewer iterations to converge.

The researchers also assessed the technique's ability to preserve dynamic contrast information. Pharmacokinetic analyses showed that ELITE maintained temporal fidelity comparable to established reconstruction approaches while enabling much finer temporal sampling. At 1-second temporal resolution, the method produced sharper arterial enhancement curves and supported quantitative measurements that distinguished healthy tissue, benign lesions and malignant lesions based on several vascular parameters.

In a blinded reader study involving two board-certified breast radiologists, ELITE demonstrated higher lesion visibility than conventional GRASP reconstructions and diagnostic sensitivity comparable to standard Cartesian VIBE imaging. While VIBE received the highest overall image quality scores, ELITE outperformed GRASP for image quality, artifact suppression and lesion conspicuity.

The authors also demonstrated the framework in head and neck and brain DCE-MRI, suggesting the reconstruction strategy could be adapted beyond breast imaging. They noted that the study was limited by its relatively small, single-center cohort and that larger multicenter studies will be needed to establish the method's clinical utility.

Back to HCB News

Artificial Intelligence Homepage