Data-Efficient Deep Learning for Kidney Segmentation on Non-Contrast MRI

Best semi-supervised and transfer learning methods achieved Dice 0.89 vs fully supervised 0.92 for kidney segmentation on non-contrast MRI, substantially reducing need for manual annotations.

• Seven frugal learning strategies were systematically compared across three datasets including a local clinical cohort and two public datasets.
• These data-efficient techniques may accelerate adoption of automated kidney segmentation in settings with limited annotated images.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.