Frugal Learning Methods for Kidney Segmentation in Non-Contrast MRI

Area of Science:

• Medical imaging
• Deep learning
• Renal imaging

Background:

• Chronic kidney disease (CKD) is a global health issue requiring early detection tools.
• Non-contrast T1-weighted MRI is a non-invasive method for kidney morphology assessment.
• Automated kidney segmentation in MRI is hindered by data scarcity and image variability.

Purpose of the Study:

• To develop and evaluate frugal learning methods for kidney segmentation in non-contrast MRI.
• To enhance segmentation accuracy with limited annotated data using U-Net architecture.
• To compare various data-efficient strategies against a fully supervised baseline.

Main Methods:

• Utilized three datasets: Barlicki (local), AMOS22, and AbdomenCT.
• Implemented seven frugal learning strategies: data augmentation, semi-supervised learning, and weak supervision.
• Compared frugal methods against a fully supervised U-Net baseline.

Main Results:

• Frugal learning methods achieved accurate and reliable kidney segmentation.
• The best semi-supervised and transfer learning models reached a Dice similarity coefficient of 0.89.
• Performance was comparable to the fully supervised model (Dice = 0.92), with reduced annotation needs.

Conclusions:

• Data-efficient deep learning techniques can overcome annotation limitations in medical imaging.
• Frugal learning accelerates the clinical adoption of automated kidney segmentation.
• These methods are particularly valuable in resource-limited settings for CKD monitoring.