Multi-Task 3D Model for Lung Nodule Feature Extraction and Classification Using CNN and Transformer

Lung cancer early detection and therapy rely on an accurate diagnosis of lung nodules. Traditional computeraided diagnosis systems(CAD) typically treat nodule malignancy classification as a single-task problem, neglecting the intrinsic correlations among multidimensional nodule features. The novel three-dimensional deep learning framework known as MT-TCNet is proposed in this research, which integrates the benefits of 3D Swin Transformer and 3D Convolutional Neural Networks (CNNs). through adaptive feature fusion and hierarchical multi-task learning structures to simultaneously predict multiple key features of lung nodules (malignancy, sphericity, margin, and subtlety). Experiments conducted on the enhanced LUNA16 dataset show that for the four classification tasks, our model achieves accuracy of $\mathbf{9 2. 6 1 \%, ~} \mathbf{9 1. 6 3 \%, ~} \mathbf{9 2. 1 2 \%}$, and $\mathbf{9 1. 6 3 \%}$, respectively. Furthermore, the hierarchical multi-task design not only enhances model performance but also improves clinical explainability, more closely approximating the diagnostic process of radiologists. MT-TCNet offers a novel approach to comprehensive lung nodule assessment and precision medicine.