Comparative Analysis of Res Net, Mobile Net, and Efficient Net Models for Lung Nodule Detection and Classification

Abstract

Cancer is one of the most lethal diseases in the world. In a country as large as India, cancer has significantly burdened the medical infrastructure and the professionals. However, it has been proven that many forms of cancer could be treated, and the survival rate would be considerably higher if the diagnostics were performed accurately and at an earlier stage. In addition to the efforts of physicians and medical professionals, computer scientists have long contributed to the medical field by creating Computer Aided Diagnostics tools. In light of the recent strides made in the realm of Deep Learning, an international cohort of researchers has contributed to the development of a diverse array of neural models and architectures. These endeavors reflect the dynamic landscape of innovation within the field. Several aspects contribute to the effectiveness of Deep Learning computer-assisted diagnostics models; nonetheless, feature extraction plays a crucial part in defining the model's effectiveness. This research investigates the capacity of deep convolutional neural networks (DCNNs) to categorise lung cancer into three distinct groups. In this study, three cutting-edge computer vision architectures, namely ResNet50, MobileNet, and Google's EfficientNet, underwent fine-tuning for the task of classifying CT scans within the LIDC-IDRI dataset. This extensive dataset encompasses 244,527 CT scans categorized into groups denoting "nodule > or =3 mm," "nodule 3 mm," and "non-nodule > or =3 mm." The primary objective was to evaluate the efficiency of the EfficientNet model family, distinguished by substantial reductions in parameters and FLOPS, within the domain of lung nodule classification. This evaluation involved a comparative analysis against the ResNet50 and MobileNet architectures. The results distinctly demonstrated that the EfficientNet model, despite its economy in parameters, outperformed both the ResNet50 and MobileNet models. Notably, the EfficientNet model exhibited notably higher ROC AUC values across all classification categories, excelling in average AUC values for the comprehensive classification task, attaining scores of 0.922 (Micro AUC) and 0.956 (Macro AUC). These findings underscore the superior performance of EfficientNet in this critical medical imaging application.