A Comparison of Transfer Learning Techniques in Lung Cancer Nodule Detection

Lung Cancer continues to be a major global health hazard, contributing to around 1.8 million fatalities per year. This number is anticipated to increase, with projections forecasting 17 million deaths by 2030. Major risk factors include asbestos, tobacco smoke, air pollution, radon exposure, previous radiation treatment, and the family history of the disease. The recent Covid-19 pandemic has worsened the situation for lung cancer patients, making them more vulnerable to complications. In this study, two Convolutional Neural Network (CNN) models, SqueezeNet and Resnet-50, were evaluated for lung nodule classification in Computed Tomography (CT) images. These CNN models differ in architecture, depth, and feature extraction abilities. Key performance indicators such as classification sensitivity, specificity, accuracy, and computational efficiency are the main focus of the comparison. The Luna 16 database, which comprise CT images and labeled lung nodules, was used for model training and validation. The results showed that SqueezeNet outperformed ResNet-50, achieving a Train-accuracy of 88.07% and a Test-accuracy of 89.62%, while ResNet-50 achieved a Train-accuracy of 83.81% and a Test-accuracy of 86.18%. Both models demonstrated strong performance in evaluation metrics like F1-score, Precision and Recall, highlighting the effectiveness of CNN-based models in enhancing lung nodule detection.