Thrombus Localization in Acute Ischemic Stroke Using Baseline Lesion Masks
Keywords: Localization, Thrombus, Ischemic stroke, CTA
TL;DR: We propose a faster alternative to patch-based methods for thrombus localization. Our method uses problem-specific information to limit the region of interest and tackle the problem.
Abstract: Thrombus localization is valuable both in treatment selection for ischemic stroke patients and downstream tasks such as extracting thrombus characteristics. We propose the z-xy method, a potentially faster alternative to patch-based thrombus localization. z-xy consists of two networks: find-z which identifies the axial slice containing the thrombus, and find-xy that locates the thrombus in that slice.
Additionally, we use lesion masks in two ways: (1) to identify and omit the healthy hemisphere; and (2) as potential input to the networks.
We compare z-xy to a state-of-the-art patch-based method, nnDetection. Both methods are evaluated on: (1) thrombus localization and (2) a downstream task that depends on localized thrombus points - thrombus segmentation.
The z-xy method performs best without the input lesion mask, with an inference time of 30 seconds per case, an average Euclidean error of 8.2 mm, and a Dice of 0.74 for thrombus segmentation. nnDetection is two times slower than z-xy but has an error of 6.3 mm. However, this difference in Euclidean error leads to only a minor improvement in Dice for nnDetection ($0.74 \pm 0.04$ vs $0.77 \pm 0.04$). Our findings indicate that z-xy is a time-efficient alternative to patch-based methods for thrombus localization, particularly in time-sensitive clinical settings of stroke.
Submission Number: 40
Loading