Metamorphic Testing for the Medical Image Classification Model

The existing studies have applied metamorphic testing technique to testing the medical image classification models, effectively alleviating the test oracle problem and reducing the testing difficulty. However, existing methods mainly focus on constructing metamorphic relations by using general image transformation methods, without combining the knowledge characteristics of medical imaging domain, resulting in problems such as low validity of metamorphic relations. According to the above problems, this paper based on the premise of conforming to the real scenario of image diagnosis, combining the key information of medical image semantics, and constructing general metamorphic relations in this field from three dimensions: the characteristics of medical images in real environment, the regular changes of lesion stage in images and the motion artifacts produced by patients in the process of filming. The medical images classification models of COVID-19 were also selected for instance validation, and the metamorphic relations were quantitatively analyzed to detect inconsistency in the classification results of different models and to assess the robustness of the model. The experimental results show that the constructed metamorphic relations by the key information of medical image semantics are able to detect inconsistencies in the models with a high detection capability, with the inconsistency percentage reaching up to 38.05%. This method can also be extended to test different types of medical image classification models.