Efficient T staging in nasopharyngeal carcinoma via deep Learning-Based Multi-Modal classification

Background

Accurate T staging of nasopharyngeal carcinoma (NPC) is crucial for precision therapeutic strategies. The T staging process is associated with significant challenges, including time consumption and variability among observers. This study aimed to develop an efficient, automated T staging system that supports personalized treatment and optimizes clinical workflows.

Methods

A total of 609 NPC patients were included, with 487 in the training cohort and 122 in the validation cohort. We employed a multi-modal learning framework that integrates MRI images and reports. Automatically delineated regions of interest (ROIs) served as masks. A hierarchical classification strategy (DeepTree) addressed complex staging challenges. We utilized Vision Transformer (ViT) to extract visual features and BERT to encode text features. To improve data fusion, we applied a Q-Former to integrate visual and textual information. The performances of the methods were evaluated using accuracy (ACC), the area under the receiver operating characteristic curve (AUC), precision, sensitivity (SEN), and specificity (SPE).

Results

Integrating images and text via Q-Former demonstrated superior performance overall, significantly surpassing single-modality methods. IT-DTM-BLIP2 demonstrated strong performance, achieving an accuracy (ACC) of 0.787 (95% CI 0.714–0.860). The area under the receiver operating characteristic curve (AUC) values were AUC1 (T2 vs. T3/T4) at 0.815 (0.71–0.900) and AUC2 (T3 vs. T4) at 0.876 (0.782–0.945).

Conclusion

Our multi-modal approach consistently performs well, offering a robust automated solution that eliminates the need for manual tumor delineation. This streamlines workflows, reduces subjectivity, and offers decision-making support that may improve workflow efficiency and encourage consistency.