Evaluating preimplantation frozen donor kidney biopsies is essential to determine the quality of kidney allografts. However, this task is hindered by interobserver variability and lack of reproducibility between pathologists due to differences in expertise levels and common artifacts in frozen preparations. While deep learning aids kidney structure segmentation on permanent sections, its application to frozen biopsies remains limited, particularly for complex structures. To address this, we developed SegRenal, an AI-driven model designed to segment key kidney compartments, including glomeruli (sclerotic and non-sclerotic), arteries, and the extent of interstitial scarring. SegRenal uses three deep learning architectures-DenseNet, UNet, and ResNet-trained on a comprehensive dataset of whole slide images (WSIs) from frozen kidney biopsies. The model was evaluated on both binary and multiclass segmentation tasks, with DenseNet demonstrating superior performance. Notably, the multiclass approach improved tissue differentiation compared to binary segmentation, particularly for complex structures. In terms of quantification, the pre-trained DenseNet model demonstrated robust performance across critical kidney structures. The model achieved a recall of 99.81% for glomeruli and 94.4% for sclerotic glomeruli. For arteries, the model reached a recall of 100%, detecting all arteries, though it generated some false positives due to the similarity of other structures like arterioles. Additionally, the model performed well in quantifying IFTA, accurately predicting the extent of fibrosis across the dataset. The inclusion of data from multiple scanners enhanced the model's generalization and robustness. By automating the quantification of these structures, SegRenal improves the accuracy and efficiency of kidney biopsy assessments, supporting more informed decision-making in kidney organ transplantation.