Sie befinden Sich nicht im Netzwerk der Universität Paderborn. Der Zugriff auf elektronische Ressourcen ist gegebenenfalls nur via VPN oder Shibboleth (DFN-AAI) möglich. mehr Informationen...
Evaluation of Explainable AI Methods in CNN Classifiers of COVID-19 CT Images
Ist Teil von
IX Latin American Congress on Biomedical Engineering and XXVIII Brazilian Congress on Biomedical Engineering, p.313-323
Ort / Verlag
Cham: Springer Nature Switzerland
Link zum Volltext
Quelle
Alma/SFX Local Collection
Beschreibungen/Notizen
In computer-aided diagnosis, many solutions based on Deep Learning were developed, but few were deployed in real environments due to the lack of transparency from intelligent models to humans. Thereafter, Explainable AI (XAI) techniques were developed to evaluate image regions with prominent influence in the decision model. In this paper, we present CAM and Grad-CAM, two XAI techniques for image classification. Particularly, we have evaluated them for interpretation of COVID-19 classification of CT images using Convolutional Neural Networks. For the classification task, we have built models with MobileNetV3, VGG-16, VGG-19, and ResNet50 using transfer learning. CT images from the Large COVID-19 CT Scan Slice Dataset were used in training and test sets. This dataset is composed of more than 17,000 CT slices labeled into three classes: covid, pneumonia and normal. This work contributed by proposing a quantitative evaluation based on Jaccard Coefficient and a proposed metric of Coverage Ratio. The metrics aim to compare the XAI regions of interest to the gold standard annotated regions by specialists. We have used another database, the COVID-19 CT Scans Dataset, for XAI evaluation. This dataset contains the regions of potential COVID-19 infection, annotated by a specialist committee. For the classification task, the MobileNetV3 model had accuracy of 97.94% and F1-Score of 98.29%. We have evaluated both XAI techniques, which yielded up to 0.7164 and 0.3105 of Coverage Ratio and Jaccard Coefficient, respectively, for CAM (MobileNetV3 classifier). And up to 0.6847 and 0.2923 of Coverage Ratio and Jaccard Coefficient, respectively, for Grad-CAM (ResNet50 classifier).