Oral Dis. 2026 Jun 5. doi: 10.1111/odi.70380. Online ahead of print.
ABSTRACT
BACKGROUND: This research was designed to establish an innovative diagnostic strategy employing whole-slide imaging (WSI) technology to address the diagnostic difficulties arising from the intricate histological architecture and morphological diversity observed in oral squamous cell carcinoma (OSCC).
METHODS: We proposed a cutting-edge multi-task learning architecture (MILGDF) that combines local-global attention mechanisms with dynamic weighted fusion. This model utilizes instance-level category-specific attention to enhance feature extraction efficacy while overcoming the limitations inherent in traditional bag-level attention methods. An adaptive weighting system was incorporated to dynamically adjust the contribution of local and global features, ensuring optimal performance in diverse prediction tasks.
RESULTS: Rigorous validation on the HIDOC and TCGA-OSCC datasets revealed the exceptional predictive performance of our model. The MILGDF framework attained an AUC of 0.952 (accuracy: 0.909) on HIDOC and 0.745 (accuracy: 0.725) on TCGA-OSCC, demonstrating statistically significant superiority over existing comparative models in both staging classification and diagnostic prediction.
CONCLUSIONS: The MILGDF model is capable of effectively utilizing information from wide-field images (WSI) for the accurate diagnosis and staging of oral squamous cell carcinoma (OSCC); its performance surpasses that of existing methods, demonstrating significant potential for clinical application.
PMID:42249621 | DOI:10.1111/odi.70380
