Multi Scale Vision Transformer Framework for Lung Module Malignancy Assessment Using CT Imaging

Lung cancer has been one of the most common causes of cancer-related death among the populations globally; the process of the adequate and timely diagnosis is one of the key factors determining the survival of the patient. Computed tomography (CT) is the most common imaging technique to evaluate pulmonary nodules; however, it is not easy to differentiate between malignant and benign or normal nodules because of similarities in image presentation: morphology, texture, and intensity. Conventional computer-aided diagnostic systems based on convolutional neural networks often cannot capture long-range spatial relations and multi-scale contextual features that are needed to make accurate nodule classification. In order to overcome these failures, a Multi-Scale Vision Transformer (MS -ViT) architecture isoffered to perform the malignancy detection of lung nodules based on CT images. The structure combines multi-resolution patch embeddings with weighted fusion attention process, thus being able to represent both local fine-grained textures and global structural subtleties in an effective way. Extensive experiments on a curated dataset of malignant, benign, and normal nodules show that the MS -ViT architecture achieved a classification accuracy of 99.9, outperforming well-established CNN baselines, including VGG19, ResNet50, Inception V3, and Efficient Net. The model is shown to be converging steadily, has a low false-negative rate, and has robust generalization across all the three classes, which implies a good potential of making successful clinical decisions in the diagnosis of lung cancer.