อัลกอริทึมการแบ่งส่วนอย่างครอบคลุมสำหรับภาพถ่ายช่องปาก

Abstract

Oral images play an important role in oral diagnosis and treatment planning. Currently, many studies have used oral images to analyze oral diseases using artificial intelligence technology. However, the oral image component includes many tissues and elements that may not be important for diagnosis, such as a patient's shirt, the doctor's fingers or equipment that appears in the image, and other patient organs outside the mouth such as the nose, eyes, etc. Consequently, the oral lesion analysis results generated by the artificial intelligent system may be inaccurate and inefficient. In this research, we defined 4 research objectives: (1) Developing deep learning models for segmenting oral tissues into 8 elements (tissues). (2) Locating the oral tissue in oral images to narrow the area for consideration. (3) Testing the hypothesis if the oral cropping algorithm has an effect on the efficiency of oral tissue segmentation. (4) Testing the hypothesis if the image resizing may affect the oral tissue segmentation performance. The results indicate that each model in this study effectively segmented oral tissues, accurately found oral tissues in the oral images, and precisely perform the oral cropping, achieving the Sensitivity, Specificity, and F1-score performance of over 94%, 97%, and 94%, respectively. We believe that this study has laid an essential foundation for future research and development of artificial intelligence systems for diagnosing diseases in oral images.