An Overview and Comparison of XAI Methods for Object Detection in Computer Tomography

Abstract

Modern hardware and software developments in the medical field generate massive amounts of data that clinicians need to analyze. Many solutions based on deep learning have been introduced to support the diagnostic process. Nonetheless, the transparency and reasoning of such systems are important for medical practices that limit the application of artificial intelligence techniques that work in 'black box' scenarios. The purpose of this paper is to present algorithms that allow interpretation of the complex structure of models used in object detection. Based on the ablation study results, a detailed analysis of the advantages and disadvantages of the chosen methods has been provided. Infidelity and consistency metrics were used to assess the algorithms of explanation.