AI-driven Convolutional Neural Network for Early Detection of Oral Premalignant Disorders and Oral Squamous Cell Carcinoma using Smartphone-captured Images: A Research Protocol

Introduction: The overall prevalence in the world of 4.47% is seen in Oral Potentially Malignant Disorders (OPMD), with the maximum prevalence of 10.54% found in Asian populations with approximately 1% ability to transform into malignancy, considering the malignant transformation is comparatively less, the progression of the lesion leads to increase morbidity and mortality and will lead to obtrusive treatment for the patient. The psycho-stimulating qualities keep betel quid popular even though there is a risk of developing oral cancer. Hence, prompt detection of these OPMD is very crucial, so that the survival rate and treatment cost decrease. The need of the hour is low-cost, easy-to-use imaging devices so that an early diagnosis of OPMD and Oral Squamous Cell Carcinoma (OSCC) can be made. Need of the study: In remote areas, populations are at higher risk due to limited resources and access to healthcare infrastructure. Low-cost, user-friendly screening devices, like a smartphone-simple touchscreen that can provide sufficient collection of images with transmission of data connected to the internet. An easily accessible high-performance Convolutional Neural Network (CNN) model for clinicians can be used to detect and classify clinically potentially malignant oral lesions and precancerous neoplasms so that an early diagnosis can be made. They can be used to detect and classify clinically oral potentially malignant lesions and precancerous neoplasms in the healthcare field, with promising results. Aim: To evaluate the diagnostic reliability of Artificial Intelligence (AI)’s deep learning algorithm-based CNN algorithm for the detection of types of OPMD and OSCC lesions in smartphonebased photographic images for early diagnosis Materials and Methods: The present cross-sectional study will be conducted in the Department of Oral Medicine and Radiology, Datta Meghe Institute of Higher Education and Research (DMIHER), Sawangi, Wardha, Maharashtra, India, from January 2024 to January 2026. The minimum subjects included in the study will be 217. The analysis will evaluate diagnostic accuracy metrics, including Sensitivity (S), Specificity (E), Positive Predictive Value (PPV), Negative Predictive Value (NPV) Area Under Curve (AUC) the Receiver Operating Characteristic (ROC) Curve for CNN performance. Image segmentation outcomes will be assessed using the Dice coefficient (F1) and IoU. Comparison with human experts will involve paired t-tests or Wilcoxon tests, supported by Cohen’s Kappa for reliability. Subgroup analysis will stratify performance by OPMD subtypes and demographic variables. A p<0.05 will be considered statistically significant.