Uncertainty-guided Multitask Mean Teacher Model for Hybrid ASPECTS Scoring of Acute Ischemic Stroke

The Alberta Stroke Program Early CT Score (ASPECTS) is crucial for accurate diagnosis and treatment decision-making in acute ischemic stroke. Recent hybrid scoring methods that combine ASPECTS prediction with auxiliary lesion segmentation have demonstrated great potential by enhancing the model's discriminative power through multitask learning. However, their performance is still limited by the scarcity of labeled data and the insufficient exploitation of shared information across tasks, posing challenges for precise and robust stroke assessment. In this paper, we propose a novel semi-supervised hybrid ASPECTS scoring framework (Semi-HAS) built upon an uncertainty guided multitask mean teacher architecture, which leverages readily available unlabeled data to enhance ASPECTS scoring. Specifically, a dedicated task interaction component is introduced in Semi-HAS to explicitly extract lesion-related complementary information from the auxiliary segmentation task. The proposed component is meanwhile coordinated with a cross-task alignment loss to successfully facilitate collaboration between different tasks for more consistent ASPECTS prediction. Furthermore, we design an adaptive uncertainty estimation for both tasks to generate dual uncertainty maps for mitigating noise introduced by unlabeled data. As a result, this innovative strategy effectively guides cross-model consistency learning for transferring trustworthy knowledge from teacher to student model and improving scoring accuracy for complex anatomical conditions. Extensive experiments on both public and in-house datasets demonstrate that Semi-HAS consistently outperforms state-of-the-art methods by achieving superior scoring accuracy across ASPECTS regions and providing reliable patient level assessments, highlighting its promise in facilitating clinical ASPECTS evaluation and promoting personalized treatments.