Abstract:
In current times, there has been a multifold growth in the prevalence of lung illness, which is responsible for millions of deaths yearly. It is now absolutely necessary to have a method of diagnosing lung illness that is rapid, accurate, and reasonably priced in order to address the epidemic. The purpose of this research was to offer a multiclass categorization of lung illness based on forward chest X-ray imaging by employing a finely tuned CNN model. The categorization is based on the four different disease classes that might affect the lungs. These disease classes include COVID-19, tuberculosis, pneumonia, and normal class. The dataset is a compilation of other people's work that was obtained through the open-source website Kaggle. Following the completion of the preprocessing step, all of 7135 X-ray pictures were loaded into the model so that it could perform categorization. At the outset, the dataset was run through five different pre-trained CNN models: VGG16, VGG19, Mobile Net, MobileNetV2, and InceptionV3. After that, a CNN model known as LungNet-7 was used for training purposes. The accuracy reached by the CNN was the greatest among them, coming up at 96.07%. The fundamental structure of the LungNet-7 model was used as the basis for the construction of the in order to further increase the classification accuracy. As part of the research, an ablation study was carried out to identify the various hyper-parameters. The suggested model attained an impressively high level of accuracy of 98.83% by using the Adam Optimizer. As part of the process of validating the performance of the architecture, multiple performance matrices were also constructed.
