An Attention-Based Multi-Task Deep Learning Model for Predicting the Primary Site of Cervical Metastatic Squamous Cell Carcinoma with Unknown Primary
Synopsis
In a single-center retrospective study, an attention-based multi-task deep learning model combining a ResNet+CBAM imaging branch with a DNN+Transformer clinical-feature branch used contrast-enhanced neck CT to jointly predict lymph node malignancy and primary site, achieving in the test set (n=457) an AUC of 0.851 for benign/malignant prediction, a Micro-AUC of 0.819 for primary-site prediction, and Top-1/Top-3 accuracy of 79%/93%, and in a real-world CMSCCUP cohort (172 cases) a Micro-AUC of 0.809 and Top-3 accuracy of 88%, significantly improving junior physicians' predictive accuracy (all P<0.001).
Interpretation
It proposes and validates an attention-based multi-task deep learning model that simultaneously predicts whether cervical lymph nodes are malignant metastases and where the primary site is. Compared with prior single-task or imaging-only approaches, it jointly combines a ResNet+CBAM imaging branch with a DNN+Transformer clinical-feature branch under a multi-task framework and applies super-resolution reconstruction to enhance CT image precision. Single-center retrospective design with 1,829 training and 457 test cases; test-set benign/malignant AUC 0.851, primary-site Micro-AUC 0.819, Top-1/Top-3 of 79%/93%.
The model retains reasonable discrimination in a real-world CMSCCUP cohort and provides assistive gain for less experienced physicians. Beyond evaluating only on data with known primary sites, the study adds a 172-case real-world CMSCCUP cohort (86 patients and 86 normal controls) for validation. In this cohort the model reached a Micro-AUC of 0.809 against the gold standard and Top-3 accuracy of 88%, with improvement in junior physicians' predictive accuracy at all P<0.001.
It provides visualization evidence of the model's focus areas through gradient-weighted class activation heatmaps. Rather than reporting numerical metrics alone, the heatmaps display the lymph node levels and imaging feature regions the model emphasizes, offering interpretability cues. The text describes heatmaps that clearly delineate lymph node levels and imaging feature regions, which is qualitative visualization evidence.
Perspective
The results target clinical assistance settings that use contrast-enhanced neck CT and clinical features to determine the primary site of cervical metastatic squamous cell carcinoma, and may especially help less experienced physicians narrow the search for the primary tumor; the conclusions rest on single-center retrospective data and specific inclusion criteria, so they apply to settings similar to the studied population and imaging workflow.
Readers will still watch how the model performs across multiple centers, different scanners, and populations, and how consistent the heatmap-highlighted regions are with clinical reasoning; in addition, the loaded text is abstract-level content lacking figures and finer stratified results, so assessing subgroup performance and error sources would require the full original materials.