Computational Image Analysis of Renal Transplant Biopsies to Predict Graft Outcome

点击翻译按钮获取中文摘要

Project Summary Kidney transplantation is the most effective modality for treating end stage kidney disease. It provides superior quality of life and significantly improves survival over dialysis. However, the demand for kidney transplants has surpassed the supply of usable organs. Because of this deficit, it is important to improve the outcomes of first- time transplant recipients through intelligent management, thereby optimizing donor organ allocation and reducing the need for secondary transplants. In assessing the health of a renal allograft, time is of critical importance. Being able to precisely predict delayed graft dysfunction and modifying treatment strategies accordingly would be greatly impactful in decreasing chronic rejection events. Existing clinical methods, such as the Kidney Donor Profile Index, which are based solely on donor demographics and clinical data, are minimally to moderately predictive of allograft outcomes. Further, current visual, semi-quantitative transplant biopsy scoring metrics, e.g., Banff, the Maryland Aggregate Pathology Index, and Remuzzi are often not predictive of renal graft function. Digital image analytical methods that quantify chronic changes in kidney that cannot be done visually, may offer clues to long-term allograft outcome. Therefore, to address the unmet need of intelligent renal transplant management, we propose a comprehensive multimodal framework, integrating high-resolution renal transplant biopsy digital whole-slide images (WSIs), and donor and recipient clinical, demographic, and social determinants of health data. Using this framework we will combine computer vision and explainable artificial intelligence (XAI) tools to derive autonomous diagnostic and prognostic models for data-driven, long-term management of renal allografts. As part of their preliminary work, the investigator team has developed a computational tool to quantify interstitial fibrosis and tubular atrophy, a chronicity measure in renal transplant biopsies, and demonstrated that the prediction of estimated glomerular filtration rate at a later time-point after biopsy using machine learning (ML)-derived image features outperforms those based on routine visual assessment. This tool will be expanded to incorporate a variety of additional analyses including robust segmentation of renal compartments in WSIs, leveraging pathologist guided attention to train deep-learning models, state-of-the-art transformer models for multi-task learning, and XAI to increase interoperability and accessibility of ML-derived predictions to pathologists. The performance of this pipeline to predict renal allograft function in a future time-point will be compared with existing methods used in a clinical setting as well as ML- based methods used for explainable prediction of disease progression in other areas of digital pathology. The tool will be deployed on a cloud-based platform and the usability by important stakeholders, namely, transplant renal pathologists, nephrologists, and surgeons will be studied with a goal to eventually include the tool in clinical workflows. The proposed work will be an invaluable asset for clinicians to take advantage of large collections of renal transplant biopsy WSIs and inform treatment decisions towards improving renal allograft function.