Automatic Assessment of ASPECTS Using Semi-Supervised Learning in Acute Ischemic Stroke

The ASPECTS is a reliable method for assessing early ischemic changes in the middle cerebral artery territory of patients with acute ischemic stroke. Despite advancements in previous scoring methods, training neural networks still requires large amounts of labeled data. The annotation process for medical image is costly and time-consuming. In this paper, we propose a novel semi-supervised framework for ASPECTS scoring in acute ischemic stroke. Our method effectively utilizes unlabeled data to promote consistent predictions across various perturbations of the same DWI input. The framework comprises both a student model and a teacher model. Within the student model, precise lesion segmentation is achieved using a 3D lesion segmentation network, followed by an ASPECTS scoring module that calculates the ASPECTS score based on the overlap between the ASPECTS regions and the lesion area. The student model learns from the teacher model by minimizing segmentation and consistency losses in relation to the teacher's output. Ultimately, experimental results on the STAR dataset confirm the effectiveness of our approach.