Deep Neural Patchworks Predict Renal Imaging Biomarkers From Non-contrast MRI via Knowledge Transfer From...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Contrast-enhanced (CE) MRI offers clear renal compartment visualization but has contraindications.
• Routine non-contrast-enhanced (NCE) T1-weighted MRI is often preferred but lacks detailed corticomedullary contrast.
• Automated extraction of renal imaging biomarkers from NCE MRI is desirable.

Purpose of the Study:

• To enable automated extraction of renal imaging biomarkers from routine NCE T1-weighted MRI.
• To achieve this by transferring compartment labels derived from CE MRI.
• To validate the accuracy and reliability of this contrast-free approach.

Main Methods:

• Retrospective study of 200 patients with paired CE and NCE T1-weighted MRI.
• Manual segmentation of renal compartments (cortex, medulla, sinus) on CE MRI.
• Rigid transfer of CE-derived labels to NCE MRI for training a 3D Deep Neural Patchworks (DNP) model.
• Evaluation on an independent test set using Dice similarity and volumetric agreement metrics.

Main Results:

• Whole-kidney segmentation achieved high Dice similarity (0.950-0.953).
• Total kidney volume showed high agreement with minimal bias (CCC 0.983).
• Cortex and medulla volumes showed minor over/underestimation, respectively, with compartment fractions shifting slightly.

Conclusions:

• CE-supervised knowledge transfer enables accurate total kidney volumetry from NCE MRI.
• This method supports contrast-free renal biomarker extraction in kidneys without significant structural distortion.
• The approach facilitates routine use of NCE MRI for quantitative renal analysis.