척추 MRI의 공간 정보 활용을 위한 비지도 학습 방법 고찰

다국어 초록 (Multilingual Abstract)

This study explores effective deep learning-based pre-training methods that leverage spatial information for precise semantic segmentation of vertebral magnetic resonance imaging (MRI). Specifically, it examines and compares the Jigsaw puzzle and RotNet techniques, as well as assesses the efficacy of Masked Autoencoders (MAE) within limited medical datasets. The results confirm that the Jigsaw puzzle approach effectively acquires spatial information, achieving superior performance in semantic segmentation of vertebral MRI. Additionally, it was observed that in scenarios with extremely limited medical data, pre-training methods from the Deep Convolutional Neural Networks (DCNN) are more effective than MAE, which employs Vision Transformer, especially when compared to natural images.