X-RAY BASED AUTOMATIC DETECTION OF BRAIN COIL COMPACTION USING UNSUPERVISED LEARNING

Endovascular coiling (EC) is a vital procedure that treats intracranial aneurysms (IA) but a common complication is aneurysm recurrence as a result of coil compaction, when the implanted coil fails to isolate IA from cerebrovascular circulation. Such an event may lead to devastating hemorrhages. Hence, frequent follow-up imaging sessions using Digital Subtraction Angiography (DSA) are critical. However, DSA is invasive, expensive and not widely available. Recently, it has been shown that skull X-rays could be used as a proxy. In this work, we present a new pipeline that enables the semi-automatic evaluation of coil compaction based on X-ray images. Our pipeline involves coil segmentation with Grab-Cut and an autoencoder that learns image embeddings with a location-sensitive loss function. This approach generates efficient representations without training on image labels. We show that the image embeddings capture information relevant to coil compaction and that a simple distance measure between embeddings outperforms other baseline methods including a Siamese network.