의료영상 분할 모델의 도메인 일반화 성능 향상을 위한 자기 지도 학습의 활용

In recent years, dep learning technology has ben widely used for medical image analysis. However, dep neuralnetworks tend to produce lower generalization performance for data in novel domains, which is afrequent scenarioin the field of medical imaging since the domain can be easily shifted by a patient’s physical characteristics andimage acquisiton equipment. Meanwhile, self-supervised learning is recently known not only to further enhancethe performance of a model, but also to improve the robustnes of it. Based on this finding, we empircalydemonstrated that a model’s domain generalization performance can be improved by using self-supervisedpre-training in thistudy. Moreover, we aditonaly found that data augmentation aplied to the pretextask cansignifcantly impact on domain generalization performance of a model.