Artificial Intelligence-Based Kidney Segmentation With Modified Cycle-Consistent Generative Adversarial Network and Appearance-Based Shape Prior

This study presents an innovative deep learning framework for kidney segmentation in magnetic resonance imaging (MRI) data. The framework integrates both kidney appearance and prior shape information using a residual cycle-consistent generative adversarial network (CycleGAN). An appearance-based shape prior model is developed, utilizing iso-circular contours generated from the kidney centroid and employing the fast marching level sets method for shape extraction. By utilizing the kidney centroid and matching cross-circular iso-circular contours' appearance, the proposed appearance-based shape prior model remains invariant to translation, rotation, and scaling, eliminating the need for alignment. Additionally, a novel weighted loss function, the H-Loss, is introduced to enhance segmentation performance and prevent overfitting. The proposed approach is tested on 34 blood-oxygen-level-dependent (BOLD) grafts from patients in our kidney transplant program, achieving an average dice score of 92%. These promising results validate the effectiveness of the approach, with optimized hyperparameters ensuring high segmentation quality.