[Attention-based multi-task deep learning model for predicting the primary site of cervical metastatic squamous cell carcinoma with unknown primary].

Objective: This study aimed to construct an attention-based multi-task deep learning model that utilizes readily available CT images and clinical features to predict the primary site of cervical metastatic squamous cell carcinoma of unknown primary (CMSCCUP). Methods: In a single-center retrospective design, we enrolled 2 286 patients with cervical lymph node metastases from known primary sites and 1 264 normal controls (individuals with benign diseases, contrast-enhanced neck CT, and no malignancy history) from Qilu Hospital of Shandong University; CT images (2D maximum cross-sections of lymph nodes in levels Ⅰ-Ⅵ, with multi-channel fusion) were enhanced via super-resolution reconstruction, and a deep learning model based on ResNet with a CBAM attention module was trained. Clinical features were further integrated via a deep neural network combined with a Transformer module under a multi-task learning framework to simultaneously predict malignancy and the primary site, and gradient-weighted class activation heatmaps were used to visualize the model's focus areas. Results: The model showed good performance on the training (n=1 829) and test (n=457) sets: in the test set, the AUC for benign/malignant prediction was 0.851, the Micro-AUC for primary site prediction was 0.819, and Top-1 and Top-3 accuracy reached 79% and 93%, respectively; heatmaps clearly delineated the lymph node levels and imaging features of interest. In a real-world CMSCCUP cohort (172 cases, including 86 CMSCCUP patients and 86 normal controls), the model achieved a Micro-AUC of 0.809 against the gold standard and a Top-3 accuracy of 88%, significantly improving the predictive accuracy of junior physicians (all P<0.001). Conclusions: Collectively, the attention-based multi-task deep learning model can effectively leverage CT imaging and clinical features to predict the primary site in CMSCCUP, demonstrating considerable potential for clinical decision support.