การเพิ่มข้อมูลภาพด้วยวิธีการเรียนรู้เชิงลึกเพื่อปรับปรุงการแบ่งส่วนรอยโรคไลเคนแพลนัสในภาพถ่ายช่องปาก

Abstract

Oral cancer is a life-threatening disease. In the early stage, the symptom usually appears as an oral potentially malignant disorder (OPMD). If the lesions spread out and are left untreated, there will be a considerable chance of progressing to oral cancer. In such a case, it will be more difficult to treat, the chance of survival is dim, and the cost of treatment is much higher. Early detection of oral lesions can bring patients into appropriate treatment, which in turn reduces the mortality rate. Additionally, we found that oral public health personnel usually encounter a critical problem when performing the oral lesion screening for elderly people on site. During the operation, the personnel have to take patient’s oral images and sort them into standard views. The work is difficult because there are a large number of photographic images by the end of the day. In this study, we developed an artificial intelligence (AI) to automatically classify oral photographic images to standard views. Also, to better serve the dentist, we developed segmentation algorithms of oral lichen planus lesions, and a generative AI to synthesize realistic oral lesions for improving the performance of the segmentation algorithm. The results show that the classification and segmentation algorithms performed remarkably with the blind dataset. Moreover, the generative AI of CycleGANs and the novel blending algorithm was able to generate realistic lesions that experts could not even distinguish between the real and fake. Finally, we found that data augmenting with the synthesized images could improve the performance of the lesion segmentation. We believe this study will lay a significant foundation for the development of oral lesion screening systems for Thai people and the world.