Efficient Lung Cancer Identification through Integrated Deep Learning on Residual U-Net Segmented CT Images and Swin Transformer Features

Abstract

Late detection of lung cancer continues to be the primary reason for its high mortality rate worldwide, which is responsible for almost 85% of all cases. Computed Tomography (CT) imaging is a powerful tool for locating lung nodules and abnormalities in a short time, thus refining the medical diagnostic process. Advances in Artificial Intelligence (AI), especially Deep Learning (DL), have significantly improved the identification of lung nodules in CT imaging. This research aims to develop a diagnostic system capable of accurately predicting lung nodules. The study harnessed two major CT imaging datasets, NSCLC Radiomics and LUNA16, for pattern analysis related to lung cancer. The residual U-Net model was found to be highly effective in segmenting lung cancer regions, with 96.21% accuracy and a Dice Coefficient (Diceco) of 0.934, demonstrating its capability to accurately capture the complex features of lung cancer areas. The Swin Transformer and Principal Component Analysis (PCA) are combined to optimize feature engineering for the segmented CT scan images. The Swin Transformer generates feature vectors of a very high-dimensional space, and PCA takes these as inputs, reducing the dimensionality of the feature space and discarding redundant features to facilitate the selection of the most relevant ones. While investigating lung nodule patterns in the NSCLC radiomics dataset, the residual U-Net model in combination with DenseNet169, ResNet50, and ResNet101 models was able to achieve a very high accuracy level and thus perform better than LUNA16. The integrated residual U-Net and ResNet101 demonstrated outstanding accuracy of 98.97%, an F1 score of 96.21%, and a Diceco of 0.946, highlighting its exceptional ability to accurately detect lung nodules.