#1866 Deep learning model for predicting delayed graft function using pre-implantation kidney images: interim results from the DeepGraft project

Abstract

Abstract Background and Aims Delayed graft function (DGF) poses a significant clinical challenge in kidney transplantation due to its adverse impact on outcomes and the difficulty of prediction. Advances in machine learning and imaging technologies could be valuable for improving predictive accuracy. This study introduces an innovative approach leveraging deep learning to enhance the prediction of DGF, a step towards personalized transplant management. This study aimed to develop a predictive model for DGF using convolutional neural networks (CNNs) to analyze pre-implantation kidney images from deceased donors. Conventional machine learning models based on clinical variables were also evaluated for comparison. The primary goal is to assess the feasibility and performance of image-based deep learning models in predicting DGF. Method In this multicenter study, pre-implantation kidney images were collected from deceased donors at multiple research sites. A standardized protocol ensured the consistency of image capture, with three images per kidney taken from different angles using smartphones at a fixed distance of 30 cm. A CNN model was trained to extract visual features from these images, focusing on patterns related to perfusion and morphology. Advanced preprocessing steps, including background removal and resolution standardization, were applied to enhance image quality and improve model robustness. DGF was defined as the need for dialysis within the first week post-transplant. To benchmark performance, a clinical-data-based machine learning model trained on traditional donor and recipient variables, such as donor age, cold ischemia time, and recipient comorbidities, was also developed. Model performance was evaluated using the receiver operating characteristic area under the curve (ROC-AUC), accuracy, sensitivity, and specificity metrics. Results A total of 234 images from 78 donors were analysed. The CNN model achieved an ROC-AUC of 0.82, accuracy of 0.72 (95% CI: 0.62–0.82), sensitivity of 0.67, and specificity of 0.76, demonstrating its ability to identify subtle perfusion and morphological features predictive of DGF. The final CNN architecture consisted of four convolutional layers with max-pooling, followed by two dense layers with dropout regularization, and a binary classification output layer. In contrast, the clinical-data-based machine learning model yielded an ROC-AUC of 0.67, showing limited predictive capability compared to the image-based model. Key visual features identified by the CNN model included perfusion heterogeneity and structural irregularities, which are difficult for human observers to detect. Conclusion The deep learning model outperformed conventional approaches, identifying critical perfusion characteristics undetectable by the human eye. This underscores the transformative potential of CNN-based systems to provide accurate, non-invasive DGF prediction tools. These findings highlight the potential for integrating image-based deep learning solutions into clinical workflows, improving decision-making and transplant outcomes. Future work should expand data collection, including diverse donor and recipient populations, and validate the model's utility in broader clinical settings.

Ähnliche Arbeiten