LNMSNet: a multi-task deep learning network for pulmonary nodules segmentation and malignancy classification

Lung cancer remains the leading cause of global cancer incidence and mortality, with late-stage diagnosis contributing to a stark 5-year survival rate of only 8%. Systematic early detection via low-dose computerized tomography (LDCT) scan can dramatically improve outcomes, with survival exceeding 90% for stage I patients. Pulmonary nodules are the primary radiological precursor, but their accurate characterization is challenged by manual interpretation, inter-reader variability, and the difficulty of visually assessing small, ill-defined lesions on hundreds of CT slices.

To this end, we propose LNMSNet, which extracts Multi-Scale features from Lung Nodules for the joint segmentation and malignancy classification. The model employs a U-shaped encoder-decoder with a ResNet-18 backbone. A key innovation of our LNMSNet is the MSConv module, which uses parallel multi-scale convolutions to capture both fine-grained texture and global contextual features, thereby enlarging the receptive field and improving boundary accuracy and size invariance.

We validated the proposed LNMSNet on a multi-center external dataset of 220 CT scans from two tertiary hospitals. The model showed superior performances in both tasks compared to other multi-task models and exhibited stable generalizability across institutions.

The proposed LNMSNet effectively leverages multi-scale feature extraction and joint task optimization for accurate pulmonary nodule characterization.