The development of alternative exam method for exfoliative cytodiagnosis by deep learning

Exfoliative cytology has long been regarded as a useful screening examination in oral medicine because it is simple to perform, minimally invasive, and imposes little burden on patients. Early detection of oral mucosal disorders, particularly oral precancerous and cancerous lesions, is critical for improving treatment outcomes and survival. However, the reliability of exfoliative cytology remains limited. Diagnostic accuracy is often compromised by artifacts such as blood and saliva, as well as variability in the number and quality of collected cells, which largely depend on the operator’s technique. These limitations have hindered its widespread clinical adoption, underscoring the need for more consistent and operator-independent diagnostic approaches.

In the field of oral surgery, clinical photographs of lesions are routinely obtained using single- lens reflex cameras for medical records. Such images are readily available and reflect real- world clinical practice, but they are generally unsuitable for image analysis due to the lack of standardization, differences in lighting conditions, and, in some cases, the use of mirrors during capture. Despite these challenges, we hypothesized that advances in deep learning could enable the extraction of clinically relevant information from these non-standardized images, thereby providing a novel diagnostic pathway.

In this study, we developed a convolutional neural network (CNN)-based deep learning model to predict cytological atypia directly from general clinical photographs. The model was trained to classify cytological atypia into five categories: Class I (normal), Class II (atypical), Class III (intermediate), Class IV (suggestive of cancer), and Class V (positive cancer). The model achieved strong diagnostic performance, with an accuracy of 0.9487, recall of 0.9454, precision of 0.9566, F1-score of 0.9504, and mean AUC of 0.9950. These results demonstrate that deep learning applied to routine clinical photographs can achieve high accuracy and reliability, independent of operator technique. In conclusion, this approach may serve as a non-invasive and practical diagnostic aid for oral mucosal disorders. By leveraging images that are already collected in routine practice, it has the potential to expand access to early detection and support clinical decision-making in oral healthcare.