Utilizing Shape-features and GAN-Based Model for Renal Segmentation in DW-MRI

Medical image segmentation plays a crucial role in diagnosing several health conditions. This is particularly significant for kidney transplantation, where identifying those at risk of renal rejection is a major challenge. This paper proposes a novel framework for renal segmentation for binary masks generation from diffusion-weighted magnetic resonance imaging (DW-MRI) scans of kidney-transplant patients. The introduced framework improves the segmentation performance by applying fusion between the shape features and Generative Adversarial Networks (GANs), thus reducing the dependency on image alignment and data augmentation techniques. The approach is as follows: (i) enhancing the images' contrast by calculating the Cumulative Distribution Function, (ii) extracting the region of interest (ROI) to delineate kidney structure, (iii) generating contour-based probability maps to differentiate between kidney and non-kidney tissues, and (iv) averaging the ground truth dataset to highlight shared kidney features. Afterward, these shape features represent the RGB format of the GAN input image: the first channel is the ROI, the second channel is the probability map of the kidney region, and the last channel is the shared layer from the ground truth dataset. Finally, we employed a residual Pix2Pix GAN model to transform the RGB images into accurate binary masks. This approach was tested on a cohort of 37 patients suspected of renal rejection through performance metrics, reporting a Dice Score of 88% and an accuracy of 94%. These results underscore the effectiveness of our framework to advance the medical imaging segmentation field and offer promising solutions for both clinical applications and future research.