BO- MobileXNet for automated oral squamous cell carcinoma risk stratification: Enhancing accuracy and clinical integration
B. Aishwarya, J. Priyanka, Mummadi Shyam Kumar Reddy, R. Mohan, Divya Biligere Shivanna, Madhusudan S. Astekar, Roopa S Rao · Digital Dentistry Journal · 2025
ABSTRACT Background Oral squamous cell carcinoma (OSCC) has high mortality, especially in underdeveloped regions. Traditional Broders’ grading is subjective and lacks prognostic accuracy, ignoring key tumor features like invasion patterns and tumor-stroma interactions. While the WHO 4th and AJCC 8th editions introduced updates, their clinical utility remains limited due to inter-observer variability and the exclusion of molecular markers. Objective To enhance OSCC risk stratification by developing a Bayesian-optimized MobileXNet (BO-MobileXNet) model using Hematoxylin and Eosin (H&E) Whole Slide Imaging (WSI). This model improves classification efficiency and accuracy over conventional CNNs, with pathologists’ validation ensuring clinical relevance. Materials and Methods Digitized WSIs from OSCC archival FFPE blocks were preprocessed through normalization and augmentation. A Deep Learning (DL) model was trained and optimized using Bayesian hyperparameter tuning for OSCC risk classification. Results The BO-MobileXNet model achieved an accuracy of 89% with a 95% Confidence Interval (CI) of 76.03%–100%, demonstrating high classification performance. The AUC-ROC score of 0.995 (CI: 0.992–0.998) indicates strong sensitivity and specificity. The binomial test yielded a p-value of 0.0032 (p < 0.05), confirming a statistically significant improvement over traditional CNN models. Additionally, a user-friendly interface was developed to facilitate seamless clinical integration. Conclusion This AI-powered approach improves OSCC risk classification, accelerates diagnosis and aids clinical decision-making. The process of diagnosing has been automated which increases the efficiency and is cost effective. Despite its great accuracy, validation with larger and more diversified datasets is required.