AI in Lung Health

Lung nodules are areas of higher density in the lungs that can happen for a number of reasons, such as smoking or being exposed to airborne pollutants for a long time. It is essential to find and classify tumors on Computed Tomography (CT) scans as soon as possible so that lung diseases can be diagnosed and evaluated, as well as for planning and making treatment plans. For the diagnosis, it is essential to understand the difference between typical lung diseases like Tuberculosis, Pneumonia, and lung cancer, as all the diseases have similar symptoms initially. Initially, all the diseases have respiratory symptoms like cough, difficulty breathing, and chest pain. Pulmonary infiltrates or nodules can be observed in lung cancer, pneumonia, COVID-19, and tuberculosis, posing difficulty distinguishing between the diseases. Thus, this thesis has performed the classification of different types of diseases using X-rays by proposing a novel deep-learning framework for the multi-class classification of lung diseases, including lung cancer. The experimental results show that the Visual Geometry Group Network (VGG) 19 + Convolutional Neural Network (CNN) outperformed other existing work with 96.48% accuracy in the multiclassification of lung diseases. Moreover, once lung tumor is detected, precise localization enables healthcare practitioners to ascertain the tumor's dimensions, which is crucial for staging and devising treatment strategies. Hence, this research proposes an advanced deep learning model called the Universal Network (Unet) to accurately segment lung tumors utilizing multiple types of imaging data, specifically CT and Positron Emission Tomography (PET) scans. The intricate structures of the suggested models, which incorporate several fusion approaches such as early fusion, late fusion, dense fusion, hyperdense fusion, and hyper-dense VGG16 U-Net, are discussed in detail. The experimental results, particularly the performance of the hyper-dense VGG16 model, instill confidence in the proposed models, as it outperformed all other analyses, receiving a Dice score of 73%. Survival analysis for lung cancer patients is a crucial aspect of treatment planning and outcome prediction. Therefore, in-depth stage classification using the TNM (Tumor, Node, metastases) staging system of Lung Cancer is of utmost importance. This thesis suggests an innovative method to classify the overall stage of non-small cell lung cancer (NSCLC) by employing multimodal data, including multi-view CT images and textual clinical information. A comparative analysis of Vision Transformer (ViT) and Convolutional Neural Network (CNN) architectures, evaluating both direct classification and TNM-based approaches are proposed. The experimental results prove that the ViT-based direct model achieves superior accuracy 98.75%, improving accuracy by 8.75% over the TNM-based ViT model, while also reducing computational complexity by 66.67%. Similarly, the CNN-based direct model achieves 87% accuracy, outperforming the TNM-based CNN model by 7%, with a corresponding reduction in computational demands. The use of the proposed methods in real-time can help practitioners to detect lung cancer and predict the survival of the patient effectively.