Evaluating Explainability in Transfer Learning Models for Pulmonary Nodules Classification: A Comparative Analysis of Generalizability and Interpretability

Computerized diagnostic systems have come a long way in terms of providing credible and speedy results in the diagnosis of lung cancer, which has become one of the leading causes of death worldwide in recent years. This progress is particularly true with the advancements in models based on deep convolutional neural networks (CNNs) using computed tomography (CT) images. However, the decision-making processes of such models are less than exactly interpretable, as they are considered black boxes. This makes physicians reluctant to trust and use them.The aim of this paper is to compare several transfer models that were pre-trained on the ImageNet dataset and apply them to lung cancer diagnosis, evaluating their generalizability and robustness. This comparative study implements a number of models including MobileNetV2, EfficientNetV2-L, EfficientNet-B7, DenseNet201, VGG19, VGG16, ResNet50, Xception, NasNetLarge, and InceptionV3. The models were trained on four distinct datasets to evaluate data diversity and heterogeneity. The models’ generalization capabilities were assessed using two separate datasets: IQ-OTH/NCCD and the LDCT dataset. To enhance the models explainability and trustworthiness, the Local Interpretable Model-Agnostic Explanations (LIME) method was utilized. Among the tested models, MobileNetV2 and ResNet50 demonstrated the highest performance and stability. MobileNetV2 achieved an accuracy of 99.28%, with false positive and false negative rates of 1.23% and 0%, respectively. ResNet50 achieved an accuracy of 99.38%, with false positive and false negative rates of 0% and 1.23%, respectively.