基于心肌活检组织病理图像的心脏移植急性细胞性排斥反应智能分级和辅助监测

Heart transplantation is the most effective way to save patients with end-stage cardiac disease and improve their quality of lives. Acute cellular rejection (ACR) after heart transplantation is one of the major factors affecting the patient outcome and success of cardiac surgery. Currently, endomyocardial biopsy (EMB) is considered to be the gold standard for the diagnosis of ACR after cardiac transplantation. In clinical practice the interpretation and grading of EMB tissue specimens depend on subjective judgement, suffering time-consuming, high uncertainty, and poor reproducibility. Artificial intelligence diagnosis technology showed its advantages in pathological diagnosis. However, the high cost of labeling on large scale whole-slide digital histopathological images (WSIs) limits its application in histopathological images. We propose a method for a computer-aided diagnosis based on combination of digital pathology and artificial intelligence. A semi-automated annotation method is utilized to solve the problem of big data sample training; A precise nuclear segmentation is achieved through fusing deep learning and traditional medical image process techniques; WSI markers can be intelligently identified to discriminate different pathological grades of ACR by leveraging various feature sets extracted from multiple levels; Following the results, we investigate the quantitative correlation between the image markers, individual clinical variables, and ACR grading via a deep learning fusion network to establish a reliable, non-invasive tool in order to provide a supplementary opinion in early diagnosis and clinical monitoring for ACR.

心脏移植是挽救终末心脏病病人生命和改善其生活质量的最有效治疗方式。术后急性细胞性排斥反应（ACR）是影响患者生存期和手术成败的最主要原因之一。目前判定ACR的金标准是心内膜心肌活检术(EMB)。临床实践中对EMB组织标本的解释和分级依赖于医生的主观判读，具有耗时耗力、不确定性高、可重复性差等缺点。人工智能辅助诊断技术在病理诊断的初探显示其优势，然而大尺寸全切片数字病理图像(WSI)的标注成本高，制约其在病理图像的应用。本项目提出了一种联合数字病理和人工智能的计算机辅助诊断方法。首先通过半自动半人工辅助标注图像解决大数据样本训练问题；采用深度学习结合传统医学图像处理方法实现细胞核精准分割；提取图像多层次特征，智能识别不同病理等级ACR，发现WSI标志物；构建深度学习融合网络，挖掘图像标志物、患者临床诊断指标与ACR分级的定量相关关系，从而为早期发现和临床监测ACR提供一种可靠无创的新手段。

急性细胞性排斥反应（ACR）是影响心脏移植患者生存期和手术成败的最主要原因之一。针对心肌活检组织病理图像的病灶检测和正确分级诊断对指导临床早期治疗和提高心脏移植的远期疗效具有重要意义。本研究完成了71例样本共709张心脏移植多染色病理图像以及临床数据的采集；提出了一种基于大规模高精度的数字病理图像的全新半监督深度学习淋巴细胞检测算法；发现和验证了区分不同ACR分级的病理图像标志物。在此基础上，构建了一种智能诊断网络模型，实现了ACR的自动分级和辅助监测。此外，研发了一款基于新型深度神经网络的病理图像信息评价系统，在合作医院进行成果转化。研究结果表明该智能系统与病理医师判读结果之间的一致性较高，为临床提供一种早期检测和定量评价ACR的有效的基于大规模病理图像的计算机辅助新手段。相关研究发表了英文SCI收录论文4篇，EI收录论文3篇，中文核心及科技核心期刊论文4篇，提交了3项专利申请，授权了1项专利和2项软件著作权。