Improving Explainability of Image Classification in Scenarios with Class\n Overlap: Application to COVID-19 and Pneumonia

Abstract

Trust in predictions made by machine learning models is increased if the\nmodel generalizes well on previously unseen samples and when inference is\naccompanied by cogent explanations of the reasoning behind predictions. In the\nimage classification domain, generalization can be assessed through accuracy,\nsensitivity, and specificity. Explainability can be assessed by how well the\nmodel localizes the object of interest within an image. However, both\ngeneralization and explainability through localization are degraded in\nscenarios with significant overlap between classes. We propose a method based\non binary expert networks that enhances the explainability of image\nclassifications through better localization by mitigating the model uncertainty\ninduced by class overlap. Our technique performs discriminative localization on\nimages that contain features with significant class overlap, without explicitly\ntraining for localization. Our method is particularly promising in real-world\nclass overlap scenarios, such as COVID-19 and pneumonia, where expertly labeled\ndata for localization is not readily available. This can be useful for early,\nrapid, and trustworthy screening for COVID-19.\n

Ähnliche Arbeiten