DESIGN AND IMPLEMENTATION OF DEEP LEARNING-BASED MODEL FOR LUNG DISEASE DETECTION AND CLASSIFICATION

Abstract

The limited availability of sufficiently labeled medical images poses a major challenge in developing reliable deep learning systems for lung disease diagnosis. The novel idea is to train popular pretrained convolutional neural network (CNN) models, such as VGG16, ResNet50, and DenseNet121, on subsets of different sizes of chest X-ray datasets and develop a multiclass classification model. The dataset includes images of five different lung diseases, i.e., lung tumor, bacterial pneumonia, viral pneumonia, normal, and COVID, sourced from Kaggle. Rotation, flipping, and brightness adjustment augmentation methods helped in creating a balanced dataset of 20,000 CXR images, each class having around 4000 images. This study demonstrates that the impact of image augmentation methods (balanced dataset), task-specific training of pre-trained models with a custom classification head, and optimization methods helped in achieving optimized outcomes. To ensure the outcomes are reliable, the experiments are done three times, and all models are trained on the same hyper-parameters.  The advancement in accuracy from 73.6% (500 images) to 91.41% (20,000 images) exemplifies scalability for data-limited settings. This mathematically grounded study offers suggestions for model preference in resource-constrained healthcare, enabling deployment in low-resource clinics.