Grading Classification of Tongue Cancer in Oral Images Using CNN Transfer Learning

Abstract

- Oral cancer is one of the top 10 diseases in cancer incidence worldwide and its prognosis is not good as early diagnosis is not made and so the survival rate from it has not been improved greatly. Oral cancer occurs most commonly in tongue, floor of mouth, and lower lip and 5-year survival rate is very low at 50% and if diagnosed early, the average time of survival becomes longer. Therefore, to improve the survival rate of oral cancer patients, early diagnosis and discovery and patient habit improvement including smoking are very important. Oral cancer can be diagnosed with imaging equipment like CT and MRI or through invasive biopsy, but it is not easy to make an approach and generally difficult to discover early. Therefore, to identify if early diagnosis would be possible using the intraoral images obtained through ordinary camera, the reliability of tongue cancer grade classification was evaluated by applying deep learning analysis. The collected images of oral cavity were classified into four groups: normal, inflammatory lesion, pre-cancerous lesion, and malignant tumor depending on clinical importance and three (3) types of image pretreatment were applied. In case of deep neural network, Inception-ResNet-v2 was used and 70%(1,307 images) of the pretreated images were used as training data and after image enhancement processing, transfer-learned with SGDM technique. As test data, the remaining 30% data (561 images) were used. In confusion matrix, it appeared that sensitivity was 90.37%, specificity 95.03%, accuracy 93.21%, positive predictability 92.06%, and negative predictability 93.93%, and in ROC curve, AUC was normal because normal was 0.9781, inflammatory lesion 0.9102, precancerous lesion 0.8944, and malignant tumor 0.9688, and the classification performance of inflammatory lesion or pre-cancerous lesion was inferior to that of malignant tumor. The accuracy depending on pretreatment did not show a statistically significant difference in HE-1 and HE-2 compared to standard image. The performance of the diagnostic classifier applying deep learning into intraoral images showed accuracy at the average rate of 82.9% and so showed more excellent performance than the precedent studies using different RGB images. Deep learning study using such RGB images tends to be rapidly commercialized as it has disclosed big data whose images were easy to obtain and that were preclassified and so noninvasive data can be easily obtained unlike biopsy or blood test. Intraoral image does not have preclassified and disclosed big data yet, but have a distortion in an artifact or lesion in obtaining an image and so if we can standardize the image acquisition and organize big data through correct preclassification in cooperation with clinicians although this is limited, it will be helpful for early diagnosis of tongue cancer.