Improving Early Detection and Classification of Lung Diseases With Innovative MobileNetV2 Framework

Abstract

Any condition that damage or impedes the normal operation of the lungs is classified as a lung disease, and failure to identify and address it early on can potentially lead to false outcomes. To address this challenge, two innovative techniques are proposed for lung disease classification, supporting medical professionals to diagnose and provide preventive measures at an early stage. The Proposed Model 1 integrates a custom MobileNetV2L2 architecture, that builds upon the MobileNetV2 framework through fine-tuning and customization. This model incorporates a ridge or L2 regularizer within its dense layer to enhance its performance. The Proposed Model 2, custom CNN2 built on CNN as its foundational block, is fine-tuned with ELU as the activation function, replacing ReLU, and incorporates the ridge or L2 regularization technique. The proposed research utilizes two publicly available datasets: DS1(Data Set1), which is the Lung Disease 5-class dataset, and DS2(Data Set2), which is the Lung Disease 4-class dataset and are collected from Kaggle. The results from the Proposed Model 1 provide better performance than state-of-the-art techniques like EfficientNet B0, InceptionV3, ResNet, and InceptionResNetV2. It achieved a training accuracy of 99.53%, validation accuracy of 100%, and test accuracy of 95.51%. The proposed Model 2 provides outstanding performance, with a training accuracy of 96.79%, validation accuracy of 91.56%, and testing accuracy reaching 99.26% The proposed research serves as a valuable tool for Pulmonologists, providing a secondary opinion in the diagnostic process.