Background
Oral squamous cell carcinoma (OSCC) constitutes a major oncologic challenge in Low- and Middle-Income Countries (LMICs), where late-stage diagnosis drives disproportionately high mortality rates. Early detection of oral potentially malignant disorders (OPMDs) through effective screening is critical to improving five-year survival and aligns with Sustainable Development Goal (SDG) 3.4 (reducing premature NCD mortality). However, LMIC healthcare systems are constrained by limited specialist availability and the subjective nature of conventional visual oral examinations, creating a significant diagnostic gap. Artificial intelligence (AI), particularly deep learning, offers a transformative solution by enabling objective, scalable, and accessible screening, thereby advancing SDG 3.8 (universal health coverage) through technology democratization.
Objectives
This study aimed to develop, validate, and preliminarily implement a clinically relevant AI framework for the automated classification of oral lesions into normal mucosa, OPMDs, and OSCC using routinely acquired clinical photographs. The research prioritized LMIC applicability through an equitable, capacity-building partnership model, explainable AI for clinical trust, and offline functionality for low-connectivity settings—thereby directly addressing SDG 10 (reduced inequalities) in healthcare delivery.
Methods
Conducted as an LMIC-led collaboration between Manipal College of Dental Sciences, MAHE, and Sinhgad Dental College and Hospital, Pune, this study embedded local clinical expertise at every stage—from protocol design and ethical oversight to dataset curation and validation. The methodology integrated evidence synthesis with technical development: first, a scoping review followed by an umbrella review mapped the current AI landscape in oral oncology, identifying critical gaps in generalizability and bias mitigation. Subsequently, a carefully curated and balanced dataset of 1,500 high-quality clinical images (500 per diagnostic class) was annotated by board-certified oral medicine specialists. A DenseNet-161 architecture, pre-trained on ImageNet, was fine-tuned and its performance rigorously evaluated on a held-out test set using comprehensive metrics including AUC, sensitivity, specificity, and F1-score. Model interpretability was ensured through Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize decision-making regions.
Results
The developed framework demonstrated excellent diagnostic performance, achieving a macro-average AUC of 0.94 on the independent test set (n=450). It showed particularly high sensitivity for OSCC (92.5%) and OPMDs (87.1%), with an overall balanced accuracy of 88.2% and an F1-score of 0.87. Grad-CAM visualizations confirmed that the model’s attention aligned with clinically suspicious features—such as irregular borders, erythroplakic components, and surface texture changes—providing transparent and actionable insights for clinicians. This interpretability is a cornerstone for fostering trust and facilitating integration into routine practice.
Conclusions/Implications
This research validates a high-accuracy, interpretable deep learning tool tailored for early oral cancer detection in LMIC settings. The framework’s strong sensitivity for malignant and pre-malignant lesions, combined with its explainable design, addresses key barriers to AI adoption in global health. The equitable, capacity-building partnership model (supporting SDG 17) ensures local ownership and enhances sustainability. These findings establish a solid foundation for the next phase of multi-centric, prospective clinical trials aimed at evaluating real-world impact on screening efficiency and patient outcomes. Ultimately, this work contributes a scalable, evidence-based solution toward reducing oral cancer disparities and advancing equitable cancer control in underserved populations worldwide.