包虫病是一种人畜共患的寄生虫疾病，严重危害人民健康和经济发展，其多发于肝脏，主要分为两类，肝囊型和泡型包虫病。肝包虫病在诊断中主要有以下临床问题：1）缺少一个高效、快速、诊断率高的方法。2）高发区医疗资源不均衡情况严重，导致诊断不及时。3）类型多变，基于影像的鉴别诊断准确率不高。因此，需要一个快速、高效且诊断过程标准化的方法来解决上述问题。近年来，人工智能与大数据挖掘等前沿技术应用在医疗领域成为一种趋势，深度学习在影像辅助诊断中得到广泛应用。将其应用于肝包虫病诊断中，可以缓解边远地区医疗资源匮乏的问题，挽救无数患者的生命。因此，本研究从包虫病影像检查中最常用的CT入手，基于包虫病诊断中迫切的临床问题，利用人工智能方法展开了一些工作。 首先，我们发现包虫病诊疗共识中建议的多期增强 CT扫查，存在辐射剂量大、有创等局限性，结合包虫病病理生理特点，首次提出基于CT平扫图像进行包虫病AI诊断的观点，通过研究低辐射且无创的CT图像，使患者获益。为此，研究并证实了CT平扫图像用于AI诊断的可行性，分析解决同类型研究中存在的问题，构建了数据入排标准，建立了患者数大于 900 例，标注大于20000 张的CT平扫数据集，解决了同类研究样本量少，临床代表性差的问题。基于该数据集，进行了首个放射诊断医生对包虫病的临床诊断实验，为包虫病的诊断研究提供了医生诊断的真实情况及鉴别诊断的易错点，为提高临床诊断水平提供了新的思路。 其次，我们发现包虫病肝脏存在标注时间长、病灶边界模糊等的问题，利用公共数据集，通过迁移学习，构建了肝脏分割模型，最终平均 Dice 为 0.941。包虫病图像复杂且类型多变，我们发现经典分割模型存在信息冗余，通过加入注意力机制的模型，突显重要信息，最终两类包虫病灶分割的平均 Dice 分别为 0.908 和 0.807。基于包虫病面临的三个主要临床问题和同类研究中存在的“验证悖论”问题，构建了包虫病AI诊断系统，在内部及外部测试集中平均准确率为 95.5% 和 93.5%。 最后，我们发现有活性囊型包虫 CT 序列中常包含无活性图片这一临床现象，结合医生的阅片逻辑，提出一种新型的特征相融合方法对有无活性病灶进行分类， 在内部及外部测试集中 AUC 面积分别为 0.87 和 0.81。针对包虫病人工智能诊断系统中囊型包虫和肝囊肿分类不佳的问题，通过影像组学方法进行分类，在内部及外部测试集中AUC面积分别为 0.99 以上。

Echinococcosis, a zoonosis caused by parasites of Echinococcus species which mainly occurs in the liver and can be divided into two categories: hepatic cystic echinococcosis and hepatic alveolar echinococcosis. It could seriously endanger people‘s health and social-economic welfare. The main clinical problems in the diagnosis of hepatic echinococcosis are as follows: 1）Lack of an efficient, rapid and high diagnostic rate method. 2）Serious imbalance in medical resources in high prevalence areas in China, leading to untimely diagnosis. 3）The imaging-based differential diagnosis of various types of diseases is inaccurate. Therefore, a fast, efficient and standardized diagnosis process is needed to solve the above problems. In recent years, the application of cutting-edge technologies such as big data and artificial intelligence in the medical field showed great potential, and deep learning has been widely used in image-aided diagnosis. Applying it to the diagnosis of hepatic echinococcosis can alleviate the problem of lack of medical resources in remote areas and save lives for many patients. In this study, we started with CT, the most commonly used medical imaging examination for echinococcosis and carried out some work using artificial intelligence methods with a focus on the most urgent clinical needs in disease diagnosis. First, we found that the multi-phase enhanced CT, which is strongly recommended in the diagnosis and treatment consensus of echinococcosis, has some defects such as high radiation dose and invasiveness. Combining the pathophysiological characteristics of echinococcosis, we proposed the idea of the AI diagnostic study of echinococcosis based on plain CT scan images, to achieve the need for low radiation and non-invasive CT images for the benefit of patients. To this end, the feasibility of using plain CT scan images for AI diagnosis was confirmed, the problems existing in previous research were analyzed, the data inclusion and exclusion criteria were established, and the first CT plain scan dataset with over 900 patients and over 20,000 annotations was constructed, which solved the problems of small sample size and poor clinical representation in previous studies, taking into account the problems of echinococcosis in the same type of studies. Based on this data set, the first clinical diagnosis experiment of echinococcosis was carried out by radiologists, providing a new way to improve the level of clinical diagnosis by providing the real situation of clinicians‘ diagnosis and the error-prone points of differential diagnosis for the diagnostic study of echinococcosis. Second, we found that the liver with echinococcosis has problems such as long labeling time and blurred lesion boundaries, and constructed a liver segmentation model by transfer learning using public dataset, and the final average Dice was 0.941. The echinococcosis images are complex and varied in types and we noticed that the classical segmentation model has information redundancy. By adding attention mechanism, important information is emphasized in the machine learning model. Finally, the average Dice of the two types of echinococcosis segmentation is 0.908 and 0.807 respectively. Based on the three main clinical problems faced by echinococcosis and the "verification paradox" in previous studies, an artificial intelligence diagnosis system for echinococcosis was built with an average accuracy rate in internal and external test sets of 95.5% and 93.5%. Finally, we identified that the CT sequences of active cystic echinococcosis often contain inactive images. Therefore, we proposed a novel feature fusion method to classify active and inactive lesions, combined with the clinicians‘ film reading logic, with the AUC in the internal and external test sets of 0.87 and 0.81 respectively. To address the problem of poor classification of cystic echinococcosis and hepatic cyst in the artificial intelligence diagnosis system of echinococcosis, the classification was performed by imaging histology method, and the AUC in the internal and external test sets were above 0.99, respectively.