Use of Artificial Intelligence’s deep learning algorithm based Convolutional Neural Network for automatic types and detection of Oral Potentially Malignant Disorders and Oral Squamous Cell Carcinoma in smartphone based photographic images for early diagnosis A cross sectional study

A handheld smartphone (Iphone 14 pro) will be used to collect oral images Steps to be followed: 1. The camera grid of the handheld smartphone iPhone 14 Pro settings will be used to aid in locating the centre of the lesion. The faint grid keeps the area of lesion to be used as the fixed region of interest (ROI) as the grid is divided into nine rectangles of equal size. The camera grid helps the lesion to be placed at the centre for obtaining accurate and appropriate image. This will aid in removing any irrelevant backgrounds. 2. The images will be taken of only those patients who have given informed and approved consent. All methods will be carried out in accordance with relevant guidelines and regulations. 3. The appropriate image will be captured by keeping the smartphone camera in a way where the lesion is in the centre of the square of the grid without any blurring or errors. 4. The image will be uploaded with diagnostic software via Wi-Fi Data collection: The subjects will be screened independently, by the Oral Medicine specialist. The medical history and image collection will be carried out in the daily outpatient clinic of the Oral Medicine and Radiology department and on patients of the hospital. Histopathologic reports may be required if a subsequent biopsy is performed. Histopathologic reports will be considered as gold standard. A patient with Oral Cancer would have undergone histopathologic examination. As proposed by the American Joint Committee on Cancer (AJCC), according to the TNM clinical staging system, stage I-IV will be Oral Squamous Cell Carcinoma cancer. The OSCC images of lip, buccal mucosa, tongue, and hard palate, upper and lower alveolar ridge will be characterized as exophytic, or endophytic or ulcerative in various areas of the oral cavity. Homogenous or low risk OPMD, may not have a histopathologic report. Thus, based on clinical features, the oral images will be under 4 categories i.e., normal, homogenous OPMD, non-homogenous OPMD, and OSCC. A white patch or a lesion, uniformly flat and thin in nature, and have a smooth or fissured surface with no atrophic or erosive lesions will be considered as homogeneous OPMD. Non-homogenous OPMD will be mixed red and white lesions which are atrophied or with irregular surface texture. A nodular lesion from the mucosal surface or a deep ulcer with a rough surface, everted both with unclear borders will be considered as OSCC. 3. The normal healthy mucosa presents with homogenous, neither white nor red patches. Each lesion will have one picture. If there are multiple lesions in a patient then those many images will be taken. In a healthy individual, numerous oral images can be taken which belong to different anatomic sites of oral cavity like buccal mucosa, labial mucosa, tongue, palate and floor of the mouth. One oral medicine specialist will take part in this research project to ensure appropriate annotation and it will be quality reviewed by an expert. These images/ datasets obtained will be used to train CNN. Finally, CNN will be able to predict the oral squamous cell carcinoma and oral potentially malignant disorders via an input image. Following steps WILL BE performed: 1. Preprocessing: The central area is obtained by cropping the irrelevant structures from the image so as to obtain the useful part of the image. 2. All photographic images were uploaded to the web application for image annotation. 3. Due to manual segmentation, the image used for CNN training, validation, and testing will be the largest area of intersection annotations 4. CNN models will be used in order to classify and detect lesions in the oral photographic images 5. Some of the oral images will be randomly assigned as training data for OPMD, OSCC and normal oral mucosa. Some of the images will be used as validation and will be used as testing data to confirm the accuracy