Generative-AI based system for accurate prediction of deceased donor liver-transplant (DDLT) outcome and viability

项目总结 美国有450万慢性肝病、急性肝功能衰竭和肝癌患者。在末尾 在不同阶段，患者需要接受已故供者肝移植(DDLT)，即整个肝脏被移植 来自临床上已经死亡的捐献者。在肝脏移植后，组织被快速冷冻，并从捐赠者身上提取一小部分样本 对器官进行切片和处理，以进行快速的组织病理学评估。评估必须在以下时间完成 移植器官能保持存活的时间非常有限。组织病理学的程序 “冰冻切片”的评估不同于标准的“石蜡包埋”评估，这是非常耗时的。 到从组织中提取水和其他物质的额外步骤。这些步骤限制了组织可以 用染料染色产生特殊的染色，用于纤维化和结缔组织的组织病理学评估。 马森氏三色镜(MT)在DDLT期间不适用于病理学家。因此，组织病理学评价 冷冻切片具有很高的挑战性，这可能会导致不良的结果和对可存活的边缘肝脏的排斥， 可以移植到病人身上。为了克服这些限制，这项提议的目标是 开发和验证一个基于人工智能的生成性系统，该系统可以生成硅胶中的“虚拟”幻灯片以帮助 DDLT期间的病理学家和外科医生。拟议的系统还将包括一个预测-人工智能模型 DDLT结果。生成模块将使用在纹理模式上训练的转换模型 苏木精-伊红(HE)切片中的纤维化和结缔组织。预测模块将i) 从生成的MT和输入HE幻灯片中提取特征，ii)将它们与接受者的临床数据融合，以及iii) 使用基于深度学习的模型来预测移植后的结果。拟议的系统可以极大地帮助 器官捐献组织在选择DDLT受者的过程中，同时提高DDLT的存活率。 我们的初步研究表明，所提出的系统可以产生能够进行诊断的虚拟幻灯片 都是由经验丰富的病理学家验证的。此外，研究表明，我们的预测模型可以执行 自动量化虚拟玻片中的纤维化，准确率为86%。在本提案(1)中，我们将 开发一个基于深度学习的软件，可以在数字高考幻灯片上生成增强的MT层 移植肝的冰冻切片，以及(2)我们将扩展该软件，使其能够预测 通过学习受者的组织病理学和其他临床标记物来预测肝移植后的 移植的可行性。 我们的技术的影响是：a)有效和准确的同种异体肝组织病理学评估 DDLT，b)改善移植后结果和存活率，c)降低组织学手术成本 实验室，以及d)加快病理工作流程和数字病理转型。

PROJECT SUMMARY There are 4.5 million US patients with chronic liver disease, acute liver failures, and liver cancer. At end stages, patients need to undergo deceased donor liver transplant (DDLT), where the whole liver is explanted from clinically dead donors. After liver-explant, the tissue is rapidly frozen, and a small sample from the donor organ is sliced and processed for rapid histopathological evaluation. Evaluation must be completed during the highly limited time the organ can remain viable for transplantation. The procedure of histopathological evaluation using “frozen sections” differs from standard “paraffin-embedded” one, which is time-consuming due to additional steps to extract water and other substances from tissue. Those steps limit how tissue can be stained with dyes to produce special-stains for histopathological evaluation for fibrosis and connective tissue. Masson’s Trichrome (MT) is not available to pathologists during DDLT. Thus, histopathological evaluation of frozen sections is highly challenging, which can lead to poor outcomes and rejection of viable marginal livers, which could be transplanted in patients. To overcome these limitations, the objective of this proposal is to develop and validate a Generative AI-based system that can produce in-silico “virtual” slides to assist pathologists and surgeons during DDLT. The proposed system will also include a Predictive-AI model for DDLT outcome. The generative module will use a transformation model trained on the texture patterns of fibrosis and connective tissue in input hematoxylin and eosin (HE) sections. The predictive module will i) extract features from both generated MT and input HE slides, ii) fuse them with recipient clinical data, and iii) use deep-learning based model to predict post-transplant outcome. The proposed system can significantly help organ donor organizations in the process of DDLT recipients’ selection while enhancing DDLT viability rates. Our preliminary study shows that the proposed system can produce diagnostic-enabled virtual slides that are validated by experienced pathologists. Also, the study shows our predictive model can perform automatic quantification of fibrosis in the virtual slides with accuracy of 86%. In this proposal (1) we will develop a deep learning-based software that can produce augmented MT layers on the digital HE slides of liver allograft’s frozen sections, and (2) we will extend the software to be capable of predicting the outcome of liver-transplant by learning histopathological and other clinical markers from recipients that predict post- transplant viability. The impacts of our technology are: a) efficient and accurate histopathological evaluation of liver allograft during DDLT, b) improved post-transplant outcome and survival rate, c) reduced operational cost in histology laboratories, and d) accelerated pathology workflows and digital pathology transformation.