Diagnosis of Tongue Cancer Using Medical Images of the Oral Cavity

Currently, several studies focus on the utilization of medical imaging for diagnosing oral lesions. However, the models employed in these investigations are either inaccessible or concentrate on diagnosing different oral conditions, such as potentially malignant lesions, or necessitate specific image inputs, such as those obtained via endoscopy or digital dental cameras. Consequently, there exists a requirement for a diagnostic framework specifically targeting tongue cancer. This study endeavors to construct and train an oral image classification model tailored for diagnosing tongue cancer, with a quality surpassing that of a general practitioner. In contrast to similar endeavors, we amassed the most extensive dataset for training purposes, comprising images of tongues captured from various perspectives, thereby enhancing the model's generalization capabilities. Additionally, we introduced a novel classifier model architecture, facilitating the extraction of image features with a relatively limited dataset. This architecture involves image segmentation before classification, isolating the tongue area for analysis. Thus, the novelty of this research lies in employing an original dataset, coupled with the development of criteria for oral cavity photograph quality and angles specifically for diagnosing tongue cancer, alongside specialized artificial intelligence training approaches. The resultant model exhibited commendable performance metrics: 95% accuracy, 87% specificity, 98% sensitivity, and ROC AUC of 0.97, comparable to oncologist diagnosis quality and superior to that of a general practitioner. The developed medical image classification system could serve as an initial diagnostic tool, rapidly assessing tongue condition and recommending specialist consultation if malignancy is suspected. Moreover, it holds potential for facilitating earlier disease detection, thereby improving survival prognosis and streamlining clinical screening protocols. Furthermore, our model harbors prospects for advancement. Firstly, efforts could be directed towards minimizing false positive rates. Secondly, alternative classifier architectures could be explored contingent upon enhanced computational resources. Thirdly, expansion of the dataset to encompass diverse lesions, such as inflammatory processes, benign neoplasms, and precancerous conditions, could enable the classifier to evolve into a multiclass system, albeit necessitating substantial input from medical professionals.