Multi-class Lung Disease Classification Using Convolutional Neural Networks

Lung diseases place a significant and growing burden on global healthcare systems. However, any technological tool that enables rapid and highly accurate screening of lung diseases can be crucially beneficial to healthcare professionals. Deep learning algorithms are critical in disease identification and medical decision-making, analyzing vast datasets, and uncovering patterns beyond human perception. This study aims to comprehensively evaluate artificial intelligence-based segmentation algorithms and their efficacy in segmenting X ray images, three public Convolution Neural Networks (CNN)-based semantic segmentation models: Fully Convolutional Network (FCN) using pre-trained VGG16, (U-Net) trained from scratch, and (LinkNet) provided by Keras, the lung area was segmented and then fed to Convolutional Neural Networks to classify the CRX into four classes of lung diseases. Five different pre-trained CNN architectures (Xception, VGG16, ResNet50V2, MobileNetV2, and DenseNet201) were investigated on the plain and segmented lung CXR images. An image pre processing pipeline was used in this research to enhance the classification process. The results of our experiments show that LinkNet by Keras outperformed the other segmentation networks in terms of accuracy, Jaccard, Dice, and Mean IoU of 98.13, 92.64, 96.09, and 96.49 respectively. The classification results show that classification performance from plain Chest X ray images is better than the segmented; theDenesNet201 obtained the best classification results of non-segmented images with accuracy, F1-Score 91.32, 91.30, respectively.