Research And Implementation Of Interpretable AI Diagnostic System For Thyroid Nodules

The thyroid is the largest endocrine gland in the human body,and thyroid nodules are masses inside the thyroid,which can be divided into benign and malignant.The incidence of thyroid nodule in China is among the highest in the world,which is a high incidence disease with a large number of patients.Ultrasonography has become the main method for screening and diagnosing thyroid nodules due to its high efficiency and non-invasive characteristics.Thyroid ultrasound reports usually describe the morphology and blood flow characteristics of thyroid glands,focal lesions,and lymph nodes.These characteristics are the main basis for clinicians to diagnose the nodules.Therefore,combining artificial intelligence technology and thyroid ultrasound features to establish an AI diagnostic model of thyroid nodules can assist clinicians in diagnosis and help alleviate the current situation of insufficient medical resources.At present,many scholars have invested in the research of how to apply artificial intelligence technology to the diagnosis of thyroid nodules,and some have achieved good results.However,most AI diagnosis models are only in the research stage and cannot be widely used in hospitals.The main reason is that deep learning models are inexplicable.Under the premise of low fault tolerance of medical diagnosis,people cannot trust a black box model’s judgment.Therefore,having both diagnostic capabilities and interpretation capabilities is a necessary condition for popularizing the use of AI diagnostic models.In view of the above situation,this paper studies and implements an interpretable AI diagnostic system for thyroid nodules.The system is mainly divided into two modules,namely the training module of thyroid nodule diagnosis model based on structural ultrasound characteristics and the local post-interpretation module of the diagnosis model.The model trained by the former is used to predict the ultrasound report and obtain the diagnosis result;the latter is used to analyze the behavior of the prediction model within the scope of the instance and obtain the interpretation result.The research content of this paper mainly includes the following three aspects:1)This paper describes the overall framework of thyroid nodule explainable AI diagnosis system,and analyzes the system module composition and functions of each module.On this basis,the paper explains the necessity of the thyroid nodule explainable AI diagnosis system to be composed of thyroid nodule diagnosis model training module and local post-explanation module,and describes the process of each module in detail.2)This paper constructed a thyroid nodule diagnosis model based on structural ultrasound characteristics.Firstly,the characteristics of thyroid ultrasonic text report are analyzed,and then the text report is converted into structured data by structural scanning method driven by thyroid semantic tree.Then,the label corresponding to the ultrasonic report was extracted from the pathological examination results of the patients within the appropriate time range,and the ultrasonic characteristic data set was prepared.On this basis,the Deep FM based thyroid nodule benign and malignant prediction model was trained.Finally,the structural results and the performance of thyroid nodule diagnosis model were demonstrated in the experimental part.3)The local post-interpretation method for thyroid nodule diagnosis model is introduced in detail.This method improves the drawback of the mainstream post-interpretation method LIME,which has unstable interpretation results.Specifically,the improvement is made from three points.First,when generating the disturbed data set,LIME samples each dimension independently,which causes the generated samples to deviate from the actual data distribution..This method proposes to construct bayesian network and conduct correlation sampling to obtain the generated samples considering the correlation between features and then obtain the disturbed data set.Second,in view of the defect that LIME cannot distinguish the distance between the disturbed sample and the instance to be explained when the data dimension is high,this method proposes to reduce the dimension of the data based on the auto-encoder and then calculate the weight.Third,for local fitting,the linear model used by LIME has a poor fitting effect.In this method,regression decision tree is proposed to replace the local fitting model,so that it can obtain clear feature weights and improve the local fit.Finally,through the analysis of the experimental results,it is proved that the improved partial post-interpretation method has higher stability,and the diagnosis and interpretation results are displayed in the system.In summary,through the framework design of the system and the process of each module,the construction of a thyroid nodule diagnosis model based on structured ultrasound features,and the elaboration of the local post-interpretation method of the thyroid nodule diagnosis model,this article has studied and realized the thyroid nodules interpretable AI diagnostic system.