Abstract—Lung cancer is also among the most common causesof cancer death across the globe where early detection using computed tomography (CT) is of great help in improving sur-vival rates. Deep learning models have shown excellent results in automated detection of lung nodules, but they are black-box, which makes them less interpretable and less trustable by clinicians. This paper provides a systematic comparativeanalysis of three popular explainable artificial intelligence (XAI)methods, namely, Grad-CAM, SHAP, and LIME, on a nodulevs. non-nodule convolutional neural network of the ResNet-50.The model suggested a value of the area under curve receiveroperating characteristic (AUC) and accuracy of 0.84 and 0.7946respectively. The quality of the explanation process was measuredquantitatively in terms of Intersection over Union (IoU) andcomputational efficiency. Measurable differences in localizationbehaviour and explanation time across the XAI methods arefound as a result of experimental results. The results suggest thatthere are trade-offs existing between interpretability accuracyand computational efficiency to implement explainable deeplearning systems in clinical lung imaging systems.
Pavan Kumar Illa (Wed,) studied this question.