Kruskal-Wallis Analysis of CNN Architecture Performance on CT Image Lung Nodule Classification

Abstract

CT lung image detection and classification with a Convolutional Neural Network (CNN) is an important step toward assisting radiologists faster and more precisely. CNN comprises various architectures with different characteristics and layer structures, which influence the test results. In addition, the number of epochs used during the training process has an impact on the performance of each CNN architecture. Testing variations throughout the same number of epochs on the CNN architecture yields various results depending on the analytic parameters of accuracy, precision, sensitivity, specificity, and F1-Score. The test results still do not provide statistical information on the performance of the CNN architecture in the classification of lung nodules on CT images. This study conducted a non-parametric statistical test, the Kruskal-Wallis test, to explore if there was a significant difference in the performance of the CNN architecture in classifying lung nodules from CT scan images. The results showed that the Asymp. Sig. Value for the five unpaired variables <0.05, including accuracy 0.002, precision 0.000, sensitivity 0.000, specificity 0.001, and F1-score 0.000, meaning that the hypothesis decision is accepted. In the classification of multiclass lung cancer CT images, the CNN architectures ResNet50, EfficientNetB1, MobileNetV2, VGG16, and AlexNet have considerably different average accuracy, precision, sensitivity, specificity, and F1-scores. Accordingly, after evaluating the CNN architecture with different numbers of epochs, it is known that each CNN design requires a distinct number of epochs to achieve the best test results.