在多模態腦部影像分割上研究兩階段訓練模型與跨資料集間之領域適應

Abstract

對於放射線治療來說，自動化的腦結構分割是至關重要，因為人工去畫輪廓不只需要專業的解剖知識，整個流程也非常耗時。我們在本研究用兩階段訓練來開發多模態學影像上的多結構分割。我們所提出的模型可以一次分割多達15種腦部結構。此外，我們也在 MICCAI 2020 ABCs 競賽中做一系列的分析來驗證所提出的兩階段訓練與多模態多結構分割的效果。 另一方面，即便深度學習的模型可以稀少的醫學影像資料集上拿到亮眼的表現，應用到外部測試集時常常會有表現上的落差。於是我們研究了不同種基於特徵來做的領域適方法以解決跨不同機構資料集上的領域差異。我們的方法利用了目標域的影像並達到接近使用目標域真實標注來做訓練的表現。此外，因為在額外無標注資料集上使用領域適應，來源域上的表現也獲得到了提昇。

Automatically segmenting brain structures is important for radiation therapy since manual delineation requires anatomical knowledge and the procedure is time-consuming. We develop a model for multi-structure segmentation on multi-modal images with two-stage training. Our proposed model can segment up to 15 brain structures. We evaluate the performance on the MICCAI 2020 ABCs Challenge with a comprehensive ablation study to show the efficacy of two-stage training and multi-structure segmentation on multi-modality. On the other hand, although deep learning models have achieved outstanding performance on rare medical image dataset, models often suffer from the performance drop on the external testing dataset. We study various feature-based unsupervised domain adaptation methods to address the domain shift while crossing datasets from different institutions. Our method leverages the image information from the target domain and achieves a result close to the one trained by target domain ground truth. Furthermore, we also raise up the performance on the source domain with the help of domain adaptation on an additional unlabeled dataset.