Segmentation of infarct lesions and prognosis prediction for acute ischemic stroke using non-contrast CT scans

Abstract

Background and purpose: Ischemic stroke is the most common type of stroke and the second leading cause of global mortality. Prompt and accurate diagnosis is crucial for effective treatment. Non-contrast CT (NCCT) scans are commonly employed as the first-line imaging modality to identify the infarct lesion and affected brain areas, as well as to make prognostic predictions to guide the subsequent treatment planning. However, visual evaluation of infarct lesions in NCCT scans can be subjective and inconsistent due to reliance on expert experience.

Methods: In this study, we propose an automatic method using VB-Net with dual-channel inputs to segment acute infarct lesions (AIL) on NCCT scans and extract affected ASPECTS (Alberta Stroke Program Early CT Score) regions. Secondly, we establish a prediction model to distinguish reperfused patients from non-reperfused patients after treatment, based on multi-dimensional radiological features of baseline NCCT and stroke onset time. Thirdly, we create a prediction model estimating the infarct volume after a period of time, by combining NCCT infarct volume, radiological features, and surgical decision.

Results: The median Dice coefficient of the AIL segmentation network is 0.76. Based on this, the patient triage model has an AUC of 0.837 (95 % confidence interval [CI]: 0.734-0.941), sensitivity of 0.833 (95 % CI: 0.626-0.953). The predicted follow-up infarct volume correlates strongly with the DWI ground truth, with a Pearson correlation coefficient of 0.931.

Conclusions: Our proposed pipeline offers qualitative and quantitative assessment of infarct lesions based on NCCT scans, facilitating physicians in patient triage and prognosis prediction.

Keywords: Ischemic stroke; Neural network; Non-contrast CT; Patient triage; Prognosis prediction.