Metamorphic Testing of Classification Program for the COVID-19 Intelligent Diagnosis

Abstract

The application of machine learning classification algorithms to COVID-19 for CT images assisted diagnosis not only reduces the workload of radiologists in reviewing films, but also improves the accuracy and efficiency of the assisted diagnosis results. However the instability of such machine learning models may lead to misclassification of results, and the expected output of the models may not be available due to the lack of transparency, which make the obtaining of test oracle difficultly. Thus in this paper, the metamorphic testing technique is applied to test the intelligent diagnosis classification program of COVID-19. The metamorphic relation is constructed by analyzing the characteristics of the lesion areas in the CT images of COVID-19, and compare consistency of the follow up test cases with the original test cases, that is how the failure detection rate of the program can be verified. The experimental results show that this method can detect the inconsistency of this program and it can be extended to test intelligent diagnosis classification programs of different diseases, thus further improving the accuracy of diagnosis classification programs.