Semi-supervised Intracranial Aneurysm Segmentation with Selected Unlabeled Data

The intracranial aneurysm is a common life-threatening disease, and its rupture can lead directly to subarachnoid haemorrhage, with a mortality rate of up to one-third. Therefore, the diagnosis of intracranial aneurysms is of great significance. The widespread use of advanced imaging techniques, such as computed tomography angiography (CTA) and magnetic resonance angiography (MRA), has made it possible to diagnose intracranial aneurysms at an early stage. However, manual annotation of intracranial aneurysms is very time-consuming and labour-intensive, making it difficult to obtain a sufficient amount of labeled data. For deep learning, it is difficult to train reliable segmentation models with only a small amount of labeled data. On this basis, semi-supervised aneurysm segmentation can be used to better exploit the small amount of labeled data and abundant unlabeled data. In practice, the unlabeled data may comprise both images with or without aneurysms, yet existing semi-supervised learning methods do not filter the training data to remove the data aneurysms, which may negatively impact model training as the data is purely negative. Therefore, we propose a semi-supervised approach to intracranial aneurysm segmentation with unlabeled data selection, where negative data is excluded from model training. Specifically, we train a 2D image classification network to filter negative samples without aneurysms and then a 3D image segmentation network based on the filtered data for semi-supervised aneurysm segmentation. The proposed method was evaluated on an MRA dataset, and the results show that our method performs better than the vanilla semi-supervised learning that does not exclude negative unlabeled data.