Hybrid transformer-CNN framework for precise segmentation and classification of pulmonary nodules in CT images

Objective This study aims to develop and evaluate a new transformer-based deep learning framework for segmenting and classifying lung nodules in CT images, addressing the limitations of current CNN-based and hybrid approaches. Materials and methods The proposed framework combines the global contextual modeling ability of transformers with the spatial detail preservation of convolutional neural networks (CNNs). A Vision Transformer (ViT)-based encoder processes CT images as tokenized patches to capture long-range dependencies, while a CNN-based decoder uses skip connections for multi-resolution feature fusion. The segmentation branch predicts voxel-level masks, while the classification branch distinguishes between cancerous and non-cancerous cases. The model was trained and validated on 769 CT images using 5-fold cross-validation and tested on 96 images. Evaluation metrics included Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Hausdorff Distance (HD95), pixel accuracy (PA), classification accuracy, and Area Under the Curve (AUC). The model's performance was compared to U-Net, nnU-Net, and TransUNet to evaluate its effectiveness in segmentation and classification tasks. Results The proposed model outperformed U-Net, nnU-Net, and TransUNet in both segmentation and classification tasks. For segmentation, it achieved a DSC of 0.95, IoU of 0.89, and HD95 of 2.4 mm, with pixel accuracy reaching 97.2% on the training set and 95.8% on the test set, showing its ability to classify voxels with precision. For classification, the model achieved an accuracy of 95.88% and an AUC of 96%, outperforming nnU-Net (93% AUC) and TransUNet (91% AUC). The model demonstrated fewer misclassifications, improved precision, and better generalization, highlighting its robustness. Conclusions The proposed framework establishes a new benchmark for lung nodule segmentation and classification by combining precision and scalability. Future research will focus on multi-center datasets and multi-modal imaging to enhance its clinical utility.