Recently, our conference paper on "Automated Thrombus Segmentation in Stroke NCCT Incorporating Clinical Data" was published. The paper was written as part of the collaborative project KI-SIGS, in which we develop AI-based solutions in collaboration with the University Medical Center Schleswig-Holstein (UKSH) to assist radiologists in the interpretation of computed tomography (CT) images of stroke patients.
The paper focuses on the automated segmentation of thrombi in stroke imaging, specifically in cranial non-contrast computed tomography (NCCT). The hyperdense artery Sign (HAS) is one of the earliest indicators of an ischemic stroke. We present a deep learning-based method that can automatically segment HAS while incorporating symptomatic information. The foundation of this self-learning algorithm is a dataset consisting of over 100 stroke CT scans provided by our partner at UKSH.
We investigated the impact of incorporating information about the affected side of the body, where the stroke symptoms occur, on the segmentation. These details were included as input to the network along with the corresponding CT scan. This extension improved the Dice score of the model from 0.44 to 0.52 in the 34 test cases. In 76% of the cases, a Dice score of > 0.1 was achieved, indicating that the model was able to correctly localize the thrombus. Performance differences were observed depending on the type of occlusion: occlusions in the M1 and M2 segments of the middle cerebral artery achieved a Dice score of > 0.1 in 89% and 73% of the test cases, respectively, indicating correct thrombus localization by the algorithm. Other occlusions, such as those in more distal segments or the posterior circulation, were detected in only 25% of the test cases.
Our study not only confirms the general suitability of the model but also highlights numerous benefits it offers to physicians, including accelerated diagnosis, treatment support, and precise HAS localization. Furthermore, it contributes to the standardization of treatment, leading to more efficient and accurate patient care. This is particularly relevant in treatment decisions regarding thrombectomy and thrombolysis.
Next week, we have the opportunity to present our work in the form of a poster demonstration at this year's BVM Workshop. The annual workshop on image processing for medicine (BVM) will take place from July 2nd to 4th in Braunschweig, Germany. Since 1993, the event has served as a meeting point for image processing experts from the German-speaking region.
We are excited to discuss our work with other professionals, receive valuable feedback, and establish new connections!