An end-to-end hybrid 3D CNN–transformer architecture for comprehensive lung cancer assessment from chest CT scans

Interpretation of the images of the chest computed tomography (CT) in an efficient manner is important in early detection of lung cancer, but the current methods tend to deal with nodule detection, nodule segmentation, and nodule malignancy as separated problems. In the present paper, a new framework named UniLung-CT is discussed, which is a multi-task network that collaboratively detects, segments, and classifies malignancies in one end-to-end architecture. The proposed model is a combination of 3D CNN-Transformer based common encoder and cross-task based feature interaction module, which facilitates a complementary exchange of information among the tasks. UniLung-CT generates 96.4% accuracy, 95.9% precision, 95.3% recall, 95.6% F1-score, and AUC of 0.964 when evaluated on the standardized-protocol LIDC-IDRI dataset, which is persistently higher than the recent state-of-the-art models. Ablation experiments also prove the efficiency of the integration of multi-tasks, whereas the computational analysis displays the efficiency of competitive inference. To further evaluate model generalization, the proposed framework was additionally validated on the independent LNDb dataset, demonstrating consistent diagnostic performance across external clinical data.