Improving Early Lung Cancer Detection using Modified Version of Convolutional Neural Network

Lung cancer is one of the deadliest diseases globally due to the aggressive nature of the disease, leading to the destruction of healthy tissues and disruption of essential body functions. Detection and accurate categorization of lung cancer in the initial stages is important to enhance the survival rates of patients. Existing computer-aided diagnostic (CAD) systems, however, are faced with the issue of imbalanced data distribution, leading to poor generalizability, biased predictions, and inaccurate performance metrics. The current work presents a novel deep learning approach to address challenges such as class imbalance, overfitting, underfitting, and high false-negative rates. The model is trained on lung cancer imaging data from LUNA60 and LIDC-IDRI, and preprocessing is performed including grayscale conversion, resizing, normalization, and segmentation by the U-Net architecture. Feature extraction is done by a CCG-Taming Transformer, utilizing the transformer architecture to learn intricate image features. For dealing with data imbalance, a hybrid sampling technique combining SMOTE and Tomek Links is employed. Classification is done by fine-tuning the pre-trained Tokens-to-Token Vision Transformer (T2T-ViT) on the ImageNet. Misclassification errors are minimized by an iterative optimization strategy by repeatedly applying hybrid sampling methods. Precision, recall, F1-score, and area under the curve (AUC) are employed to quantify the system performance. Results indicate that the proposed technique enhances the accuracy in lung cancer detection, ultimately contributing to better treatment efficacy and patient outcomes.