COMPARATIVE ANALYSIS OF CNN VARIANTS AND VISION TRANSFORMERS FOR EARLY LUNG CANCER DETECTION

Abstract

Lung cancer continues to be a leading cause of death from cancer, and the prognosis largely relies on prompt and precise detection. Traditional methods of disease diagnosis require the interpretation of radiographs and histopathological images that require a lot of effort and suffer from observer variability. It has become crucial, therefore, to develop deep learning models in the field of computer aided diagnosis due to the ability of modern image models to extract discriminative features of interest and make decisions quickly. This thesis aims at evaluating several deep learning architectures and comparing their diagnostic capabilities for detecting early signs of lung cancer. One major goal of the thesis is to investigate the effect of deep learning model characteristics on diagnostic performance. These characteristics include model depth, parameter efficiency, feature reuse, transfer learning capability, and attention mechanisms. Architectures of CNNs and transformers will be reviewed with respect to the abovementioned characteristics, specifically Alex-Net, VGG16, ResNet50, DenseNet121, Inception Net, EfficientNetB0, Vision Transformer (ViT), and DeiT. An explanation and justification will follow for the choice of Vision Transformer over other model types. Preprocessing steps in the experimentation framework include resizing, contrast adjustment, normalization, augmentation, and splitting of the public datasets containing histopathology lung cancer The effectiveness of the algorithms is gauged using performance metrics like accuracy precision, recall, F1 score, the area under the receiver operating characteristic curve, and also confusion matrices. The results from the comparison suggest that deep networks, especially Inception-Net DenseNet121 Efficient-Net and transformer based networks, are promising approaches for lung cancer detection because of their discriminative power and generalization ability. There are other elements that get attention too, such as limitations like dataset biases, small clinical trial groups, computational costs, model interpretability, and the need for external verification before any clinical use, full stop. Overall, this thesis kind serves as an extension of the paper that…