官方翻译一下：由于不同的疾病需根据严重程度的不同采取相应的治疗措施以提升疾病的治愈率并减少医疗资源的浪费等。因此在临床诊断中仅仅诊断出病变是什么是远远不够的还需要准确识别其严重程度以采取正确的治疗方法。然而由于消化道疾病种类较多且疾病严重程度不同分级之间的外观特征较为相似即使是经验丰富的内镜医生也很难做出准确的判断且不同的内镜医生的诊断结果也会存在差异。因此设计一个能够辅助内镜医生实现准确的疾病严

Due to the different levels of severity of different diseases, corresponding treatment measures should be taken to improve the cure rate of diseases and reduce the waste of medical resources. Therefore, it is not enough to only diagnose what the lesion is in clinical diagnosis. It is also necessary to accurately identify its severity level in order to adopt the correct treatment method. However, because there are many types of digestive diseases and the appearance characteristics between different severity levels are similar, even experienced endoscopists find it difficult to make accurate judgments, and there are differences in the diagnosis results of different endoscopists. Therefore, designing a CAD algorithm that can assist endoscopists in accurately diagnosing the severity level of diseases is a very important work. However, the CAD algorithms proposed to diagnose the severity level of diseases currently rely on a large amount of labeled data, which not only has poor generalization performance, but also requires a lot of manpower and material resources to obtain labeled data of different severity levels of diseases. Therefore, designing a model that can achieve fine-grained severity level diagnosis based on a small amount of labeled data of a certain disease is the focus of current research. To solve the above problems, this chapter designs a model that can quickly adapt to a small amount of disease severity level labeled data to achieve fine-grained severity level diagnosis. The model continuously focuses on the target area in the input through a progressive manner and extracts multi-scale feature information. Then, the extracted multi-scale features are fused and mapped to a more compact metric space through a dual similarity measure module, so that the model can pay more attention to the subtle feature differences between input samples and classify them based on the distance between features, effectively avoiding reliance on a large amount of labeled data. Through a large number of experiments, this model can learn to identify similarities and differences between samples, autonomously distinguish different samples, has strong generalization and applicability, and can achieve fine-grained severity level diagnosis of different diseases based on a small amount of severity level labeled data without any additional manual annotation