PIEGUNet: A Physics-Inspired U-Net for Kidney MRI Segmentation in Chronic Kidney Disease

Abstract

Accurate kidney segmentation from abdominal magnetic resonance imaging (MRI) is critical for identifying and treating chronic kidney disease (CKD). Nevertheless, manual segmentation is time-consuming and is subject to inter-observer variability. This study introduces PIEGUNet, a physics inspired energy guided UNet framework for automatic and reliable kidney segmentation from 2D T2-weighted abdominal MRIs, which incorporates four innovative modules inspired by mathematical physics, which improves feature sharing and stability. After 100 epochs of training, PIEGUNet achieves a Dice loss of 0.1417 and a TOPSIS score of 1, outperforming 15 leading variations of UNet. Each module supports the physics-inspired design confirmed by the ablation studies. Furthermore, a paired one-tailed t-test was performed to evaluate the statistical significance of PIEGUNet's improvement. It improves kidney MRI segmentation by providing better diagnostic decision-making in biomedical imaging.