Abstract: The ultrasonographic manifestations of thyroid nodules are complex and varied, and it is difficult to judge the benign and malignant. Ultrasound examination depends on the performance of ultrasound equipment, and is closely related to the understanding and experience of ultrasound doctors. Different sonographers have different understandings of thyroid nodules in the same patient, and their report conclusions differ greatly, which brings confusion to clinical treatment. The American College of Radiology published the ACR Thyroid Imaging Reporting and Data System (ACRTI-RADS), which proposed a risk stratification approach and gave five indicators for the diagnosis of thyroid nodules: composition, echogenicity, shapes, margins and echogenic foci. The imaging features of each indicator were described and scored in the system, and the total value of the five indicators was used to determine the malignant risk of thyroid nodules and assign them to the corresponding category. The diagnostic process has become more standardized and normalized, and it is possible to use computer-assisted diagnostic classification for thyroid nodules.
The diagnosis of thyroid nodule is a multi-attribute decision making problem, which is to assign thyroid nodules to the corresponding malignant risk grade according to the imaging features. There are two main types of multi-attribute classification decision methods: One is direct classification algorithm, that is, decision-makers directly give decision parameters such as utility function, criteria weights and classification thresholds, and use these parameters for direct classification; The other is the classification based on case learning, that is, decision makers learn from the classification results of representative sets of typical cases, build a corresponding decision model to calculate the decision parameters, which are used to classify all evaluation objects. This paper combines the advantages of both approaches and proposes a two-stage multi-criteria classification method to diagnose and classify thyroid nodules. Firstly, we take the cases diagnosed by experienced doctors as case set and quantify the criteria according to ACRTI-RADS to construct a quantifiable case set. Then, the classification diagnosis process is divided into two stages. In the first stage, all cases are grouped into decision tree models based on the direct classification criteria given in ACRTI-RADS. Each type is regarded as positive class, while the other types are regarded as negative class. The classification accuracy of the models is calculated to identify the best distinguishable class, and then the cases with significant features and good discrimination are directly classified. In the second stage, a multi-attribute classification decision model is constructed. The center of each group in the case set is taken as the reference point, and the distance between cases and the central point of the group is defined. The decision objective function is to satisfy the minimum classification errors within and between groups. The constraint space is constructed with the classification distance critical value constraint and index weight constraint. Lingo is used to solve the multi-attribute classification model through the learning of typical cases, and the criteria weights and classification thresholds are calculated, so as to complete the decision classification of other complex cases.
Data from 16 Chinese electronic medical records for thyroid ultrasound diagnosis including 4 types are obtained and quantified. Manhattan distance and Euclidian distance are respectively used as case distances, and classification calculation is carried out according to the two-stage model proposed in this paper. Moreover, classification results are compared with classical classification algorithms such as logistic regression model, hierarchical vector machine model, and direct multi-attribute decision model. The experimental results show that: (1)The classification performance of using Euclidean distance as case distance is better than that of Manhattan distance. (2)The method proposed in this paper is superior to these classical classification algorithms, whose performance depends on the learning of mass data, which is difficult in medical practice. (3)The proposed method overcomes the difficulty of direct classification of thyroid nodules due to the complexity of classification problems and the cognitive limitations of decision-makers, while taking into account the computational efficiency. In conclusion, the effectiveness of the two-stage multi-attribute classification method in the classification and diagnosis of thyroid nodules is verified by analyzing and comparing the classification results.
In the follow-up study, we will continue to collect data to improve the accuracy of the model proposed in this paper. On the other hand, this model will be extended to other medical diagnosis of multi-classification diseases, such as the diagnosis of breast nodules, the staging of hypertension, etc. In the end, thanks to the support of National Natural Science Foundation of China, which makes this research carried out smoothly, and thanks to all the experts and editors for their suggestions on modification and improvement, which enable this article to be continuously improved and successfully published.

Key words: two-stage, multi-criteria classification, case distance, thyroid nodules, intelligent diagnosis

摘要： 提出了一种两阶段多属性分类框架用于解决甲状腺结节分类诊断问题。选取经验丰富的医生诊断病例并进行指标量化,构建一个可量化计算的案例集。分类诊断过程分为两个阶段,第一阶段以ACRTI-RADS中的分类标准为依据,对案例分组构建决策树模型并识别出区分性最好的类别。第二阶段以案例分组的中心点为参考计算案例到分组中心点的距离;以组内和组间的分类误差最小为目标;以分类距离临界值和指标权重为约束构建多属性分类模型,从而对复杂案例进行分类。该方法克服了由于分类问题复杂性以及决策者认知局限性导致的对甲状腺结节直接进行临床分类诊断的困难,同时兼顾了计算效率。最后通过结果分析,并对比分类结果,验证了基于两阶段多属性分类方法在甲状腺结节分类诊断中的有效性。

关键词: 两阶段, 多属性分类, 甲状腺结节, 智能诊断