Deep Learning Approaches for Lung Cancer Classification: Assessing CNN Performance using Saliency Maps and Grad CAM on CT Imaging

Abstract

Lung cancer remains among the leading causes of mortality in the world and CT imaging is one of the crucial instruments of early diagnosis. These scans however are time consuming and prone to error in case they are manually interpreted. We use a dataset of 364 lung CT images in this study, as they have been obtained in an Iranian hospital comprising of 238 cancerous and 126 noncancerous patients, which are accorded with labels by a specialist in pulmonology. We test five different convolutional neural network designs, under the same preprocessing, augmentation and training regimes to give an equal evaluation opportunity. To solve the issue of class imbalance and low number of samples, data augmentation methods were used to increase the size of the data threefold to enhance model generalization. To make the decisions of the models more interpretable, saliency maps and Grad Cam were employed to visualize the decision-making of the models, where attention was paid to the clinically relevant areas, including tumor areas in cancerous scans and ground glass opacities in noncancerous images. The results of the experiment suggest that NASNet Mobile demonstrates excellent performance, with the highest level of test accuracy of 99.94. These results indicate that it is more effective in classifying lung cancer. In general, this paper delivers a multimodal comparison, solid data augmentation approaches, and interpretable visual features that enhance clinical confidence. It leads to the development of automated lung cancer detection, as well as offering a scalable solution that can be deployed in resource constrained settings.