Detection of Oral Diseases on Fluorescence Images Obtained Under Blue Light Irradiation

In recent years, devices for evaluating oral mucosa using light-illumination of various wavelengths have attracted attention for the detection of oral diseases. When blue light is irradiated on oral diseases, the blue-band wavelengths are absorbed and lesional tissues appear darker. However, it is difficult to accurately detect the diseased area because its appearance depends on the direction and amount of irradiation on the target area. In this study, we attempted to detect oral diseases in fluorescence images obtained under blue light irradiation using Mask R-CNN, an instance segmentation method. The success or failure of disease detection was determined by whether or not the intersection over union (IoU) with the lesioned region was greater than a set threshold value. We found that 87.0% of lesions were detected under IoU ≥ 0.5, and 97.5%under IoU > 0.0. These results indicate that the false negative rate would be very small, since a portion of the diseased area was detected in almost all cases.