Transformer-based multi-task model for lung tumor segmentation and classification in CT images

This study proposes a Transformer-based multi-task model for lung tumor segmentation and classification in CT images, enhancing accuracy and efficiency through advanced segmentation and optimized machine learning classifiers. This study utilized a dataset of 678 CT images of lung tumors, including both benign and malignant cases. Tumor segmentation was performed using three deep learning models: UNETR (Transformer-based), nnU-Net, and U-Net. The segmented tumor regions were used for radiomic feature extraction, generating 215 quantitative features, including first-order statistics, co-occurrence matrix features, and run-length matrix features. To enhance classification performance, dimensionality reduction and feature selection were applied using Principal Component Analysis (PCA), Analysis of Variance (ANOVA), and Mutual Information (MI). For classification, three machine learning models were evaluated: XGBoost, Support Vector Machine (SVM), and Stacking Ensemble. The models were trained using a five-fold cross-validation strategy to ensure robustness. Performance was assessed based on Accuracy, Sensitivity, and AUC-ROC. The proposed multi-task framework integrates segmentation and classification to enhance early detection and diagnosis of lung cancer in CT imaging. The UNETR-based segmentation model outperformed both UNet and nnU-Net, achieving higher values in key performance metrics. For example, the Dice Similarity Coefficient (DSC) was 0.96 for the UNETR, compared to 0.92 for UNet and 0.85 for nnU-Net. The Intersection over Union (IoU) was 0.95 for the UNETR, while it was 0.90 for UNet and 0.82 for nnU-Net. Additionally, the UNETR showed better results in capturing tumor boundaries, with a lower HD95 value of 2.2 mm compared to 3.2 mm for UNet and 4.3 mm for nnU-Net. In terms of classification, the Stacking Ensemble combined with PCA delivered the highest accuracy, achieving 95 % on training data and 92 % on testing data, outperforming both XGBoost and SVM. The UNETR proposed in this study outperformed traditional methods in both segmentation accuracy and classification performance.