A semi-symmetric domain adaptation network based on multi-level adversarial features for meningioma segmentation

摘要

Early diagnosis of meningiomas has led to many lives saved. And the MRI is the commonly used mean for meningiomas, which can locate the early lesions for excision to prevent growth or malignant. Due to the increasing pressure on the physicians, Computer Aided Diagnostics (CAD) systems have been widely noted to improve work efficiency. However, the available methods are mostly based on a large number of labeled samples and lack of generalization to multiple situations. Once the data source changes, the model may fail. Meanwhile the variety of imaging devices used for meningioma MRI can lead to dramatic shifts in data distribution across data sources. Considering to this the paper proposed a plug-and-play model to achieve unsupervised domain adaptation. We design a semi-symmetric structure based on partial parameters sharing for the different situations, especially for the various devices. And a novel multi-level adversarial features method is proposed to retain segmentation performance during domain adaptation. What is more, we conducted a large meningiomas dataset about different devices to verify our method’s performance and effectiveness.