DAFNet: A multi-modal deep learning model based on whole-lung CT for the automated prediction of the air space spread of lung adenocarcinoma

Spread through air spaces (STAS) is a characteristic invasive pattern of lung adenocarcinoma (LUAD), which is associated with a high recurrence rate and poor prognosis. This research introduced the automatic deep learning multimodal detection network (DAFNet) for predicting STAS based on preoperative whole-lung CT scans. In contrast to conventional approaches necessitating manual tumor delineation, DAFNet performs comprehensive end-to-end analysis of pulmonary imaging data through integrated multimodal data fusion and multiscale feature extraction methodologies. A retrospective analysis was performed on 1164 patients with LUAD (511 STAS-positive and 653 negative) from two centers, with a training-to-test split ratio of 70:30 (814 in training set and 350 in test set). In the test set, DAFNet demonstrated an area under the receiver operating characteristic curve (AUROC) of 0.90 (95% confidence interval 0.86–0.94), significantly outperforming predictive models utilizing clinical examination numerical data alone (AUROC 0.72) and radiomics features independently (AUROC 0.65). The implementation of an adaptive gate fusion mechanism combined with a DINOv3-based pre-trained architecture substantially improved predictive accuracy for STAS status determination. These findings establish DAFNet as a promising fully automated, non-invasive diagnostic tool for preoperative STAS prediction in lung adenocarcinoma, thereby advancing personalized surgical planning and promoting AI-driven oncological applications through enhanced clinical translation potential.