INTEGRATION OF AI AND RADIOMICS IN GENERAL NEPHROLOGY AND KIDNEY TRANSPLANTATION

AI and machine learning are increasingly applied across general nephrology, not only in KT. In non-transplant populations, AI-driven analysis of US and multiparametric MRI enables automated segmentation, quantitative assessment of perfusion, fibrosis, oxygenation, and tissue heterogeneity, and early risk stratification in CKD, DKD, and hypertensive nephrosclerosis. These approaches aim to identify subclinical injury patterns, predict disease progression, and support individualized therapeutic monitoring before irreversible functional decline occurs [13, 41].

In this context, deep-learning algorithms trained on multiparametric MRI and US datasets can automate segmentation, quantify perfusion, and estimate fibrosis. Deep-learning algorithms trained on multiparametric MRI and ultrasound datasets can automate segmentation, quantify perfusion, and estimate fibrosis. For example, the review by Zhao et al. summarizes multiple AI/machine learning imaging works in CKD [13]. Native _T_₁ mapping-based radiomics has been shown to differentiate renal fibrosis and functional decline with AUCs >0.9 in CKD populations.

AI in diabetic and hypertensive kidney disease

Machine-learning frameworks combining imaging-derived metrics (e.g. elastography stiffness, ASL perfusion) with clinical variables have shown potential for detecting diabetic kidney injury. The study by Ma et al. used ASL-based radiomics and ML models to detect early diabetic renal damage [37]. In hypertensive nephrosclerosis, ARFI elastography has demonstrated increased renal stiffness correlating with fibrosis and other parameters, laying the groundwork for radiomics/ML in this domain [29]. Collectively, these studies demonstrate that AI-assisted imaging in general nephrology can move beyond descriptive assessment toward quantitative phenotyping, enabling earlier detection of fibrosis, perfusion deficits, and progression risk in CKD populations independent of transplantation.

Building on these advances in general nephrology, AI applications have been particularly active in kidney transplantation, where longitudinal imaging datasets and defined clinical endpoints facilitate model development.

Machine learning and radiomics for US-based graft function prediction

The application of radiomics and machine learning to US imaging is a rapidly developing field in transplant nephrology. A 2022 study from Soochow University used a cohort of 233 kidney transplant recipients to develop machine learning models incorporating radiomic features extracted from gray-scale US images. These models achieved AUC values between 0.788 and 0.839 for differentiating between normal and impaired graft function, highlighting the potential of radiomics to enhance functional prediction beyond conventional imaging interpretation [63].

Building on this, a 2023 scoping review published in Transplantation cataloged 16 studies exploring radiomics and AI in kidney transplantation. These studies predominantly focused on predicting allograft rejection, fibrosis progression, or functional decline using a combination of imaging-derived features and clinical parameters. The review emphasized that while radiomics offers promising non-invasive biomarkers, the lack of standardization and small sample sizes remain major barriers to clinical implementation [64]. Nevertheless, these early efforts suggest that AI-powered radiomics could complement or even partially replace invasive biopsy in longitudinal monitoring of graft health.

Radiomics in multiparametric MRI for graft survival prediction

Recent advances have also extended radiomics and machine learning applications to multiparametric MRI in renal transplantation. A 2025 study utilized dynamic contrast-enhanced MRI with deep-learning algorithms to predict long-term graft survival in transplant recipients. This approach involved extracting high-dimensional radiomic features that quantified microvascular perfusion and parenchymal heterogeneity. When combined with patient clinical profiles, these models showed high accuracy in predicting 1-year graft outcomes, suggesting a potential role for MRI-based radiomics as a prognostic tool in kidney transplantation [65].

Though promising, these techniques are still in early developmental stages, and validation in larger, multicenter prospective trials is necessary. Importantly, this integration of imaging data with machine learning bridges quantitative imaging science with precision nephrology and could reshape future approaches to allograft risk stratification.

AI–augmented Doppler and functional US

AI is under investigation to being applied to enhance interpretation of Doppler and perfusion US in transplant imaging. A 2024 deep-learning fusion model for CKD (though not exclusively transplant recipients) improved detection of early fibrosis using multimodal US features, suggesting immediate translational potential in allograft care. Additionally, machine learning has been applied to US-derived deep learning workflows, automatically analyzing Doppler waveforms and microbubble dynamics in CEUS to quantify microvascular flow, though most systems are currently in proof-of-concept or diagnostic development stages. Given the increasing availability of Doppler image datasets, convolutional neural network-based tools targeting subclinical perfusion deficits or automated RI measurements represent a promising future direction for diagnostic augmentation.