ENHANCING LUNG NODULE DETECTION AND CLASSIFICATION USING VISION TRANSFORMERS IN MEDICAL IMAGING

Abstract

Lung cancer remains one of the leading causes of cancer-related deaths worldwide, primarily due to late-stage diagnosis and the difficulty of accurately identifying pulmonary nodules in early stages. Computed Tomography (CT) imaging plays a vital role in lung cancer screening; however, manual interpretation of CT scans is time-consuming, prone to inter-observer variability, and often affected by the subtle and highly variable nature of lung nodules. To address these challenges, this study proposes an automated lung nodule detection and classification framework based on deep learning techniques. The proposed approach integrates MedSAM based segmentation with a MobileViT based classification model to improve both accuracy and computational efficiency. Initially, lung nodules are segmented from CT images using MedSAM. The segmented nodules are then passed to a MobileViT network, which combines convolutional layers for local feature extraction with transformer-based self-attention mechanisms for capturing global contextual relationships. This hybrid design enables the model to effectively learn both fine-grained morphological features and long-range dependencies within nodule regions. The framework is evaluated on the LIDC-IDRI dataset and achieves strong performance with a training accuracy of 95.58%, validation accuracy of 92.13%, and test accuracy of 91.30%. Experimental results demonstrate that the proposed method provides stable learning behavior, reduced misclassification rates, and balanced performance across benign and malignant classes. The integration of segmentation and classification further improves robustness by focusing the model on clinically relevant regions and reducing background noise.