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.
