두부 계측 엑스선 영상을 이용한 교정 진단에서 거리 학습과 자기 지도 학습의 영향

Abstract

Deep learning has been applied to various fields, showing remarkable performance improvement. However, when published studies are applied to subtasks derived from common tasks or different domains, the performance improvement is often not as significant as expected or declines. To solve these problems, two types of representation learning, a study that skillfully handles the features obtained from the model, have been actively studied: metric learning and self-supervised learning. In this study, two experiments were conducted to check how representation learning, especially metric learning and self-supervised learning, affects medical images using cephalogram: 'orthodontic diagnosis with cephalogram' and 'the effect of self-supervised learning on orthodontic diagnosis'. In the first study, three orthodontic diagnoses were conducted: anteroposterior skeletal discrepancies (APSD: Class I, Class II, and Class III), vertical skeletal discrepancies (VSD: normo-divergent, hyper-divergent, and hypo-divergent), and vertical dental discrepancies (VDD: normal overbite, open bite, and deep bite). To avoid ‘the gray zone’ where individual diagnoses are overlapped, ArcFace was added to the existing model. Also, Group Normalization was used for stable training with small data instead of Batch Normalization. As a result, the proposed model has consistently shown good performance in internal validation and external validation. In the second study, pretext task was conducted using SimSiam, one of the self-supervised learning models, and downstream task was conducted in APSD. For comparison, randomly initialized weights and weights pre-trained on ImageNet dataset were used. As a result of the linear evaluation and fine tuning, SimSiam showed better performance in full and low data regimes and did not induce overfitting compared to training from Scratch and ImageNet. Both studies confirmed that metric learning and self-supervised learning in medical images could improve performance, extract discriminative features, and train models that are robust to data distribution and the number of data. In the future medical image artificial intelligence, research that incorporates representation learning should be conducted rather than simply evaluating performance by learning by model.