Lung Nodule Classification for Clinical Decision Support Systems: A Comparative Analysis of DenseNet-121 and Xception Architectures

The early and accurate assessment of lung nodules is an important clinical requirement worldwide, as it contributes to the timely evaluation of potential malignancy risk. In this study, the performances of DenseNet-121 and Xception architectures were compared for the automatic classification of lung nodules from computed tomography (CT) images. A total of 20,432 images were evaluated in two classes, Healthy and Lung_Nodule, and both models were trained using a transfer learning approach under the same data-splitting and preprocessing conditions. DenseNet-121 showed a slight overall advantage over Xception, achieving an Accuracy of 0.894, Balanced Accuracy of 0.895, and F1-Score of 0.894. For Xception, these values were obtained as 0.892, 0.894, and 0.892, respectively. When the clinical error profiles were examined, DenseNet-121 provided higher Specificity (0.946) and a lower FPR (0.054), whereas Xception showed a small advantage in reducing missed nodule images with a lower FNR (0.152). Grad-CAM analyses indicated that both models focused more clearly on nodule regions in correctly classified images. The findings suggest that, despite their similar overall performance, the two architectures differ in terms of their clinical error behavior.