Abstract Body: Introduction: A semi-automated method that compares voxel density with the contra-lateral hemisphere to generate ratio, or relative Non-Contrast CT (rNCCT) maps for identifying hypodensity changes was developed. In addition to being sensitive to stroke related hypodensities, these maps are also sensitive to motion artefacts and naturally occurring asymmetry in densities across hemispheres. We assessed the value of a deep-learning based model to segment and remove these artefacts and for identifying ischemic core of baseline NCCT.
Methods: We included data from 268 acute ischemic stroke patients with a large vessel occlusion from the ongoing CT perfusion to Predict Response to Recanalization in Ischemic Stroke Project 2 study. NCCT scans acquired at the primary stroke center were used to create rNCCT maps. These maps detect regions with at least 1% relative hypodensity difference compared to the contralateral region. A trained observer who had insight of arterial occlusion location manually annotated artefacts. We trained a no new UNet using the NCCT, rNCCT, and flipped NCCT images to detect artefacts from the rNCCT maps. To assess the extent to which our model falsely identified ischemic regions as artefact, we determined the overlap between the automatically segmented artefact on the rNCCT map and the manually segmented ischemic core on diffusion-weighted imaging (DWI) acquired at the comprehensive stroke center before treatment.
Results: The best performing model was the ensemble of the five cross-validation folds of 3d low- and high-resolution models based on dice similarity coefficient. Figure 1 provides an example of our model’s artefact segmentation and the processed rNCCT map after artefact removal. For the 54 patients (20% of study population) in our test set, our model achieved a median Dice similarity coefficient of 0.95 (IQR: 0.91-0.97) and a median false positive volume of 6.1 (3.2-11) ml. In the 30 patients with available DWI scans, 30% of patients had any overlap (>=1 voxel) between the segmented artefact and DWI ischemic core with a median overlap volume of 0.69 (IQR: 0.32-2.3) ml.
Conclusion: We demonstrate the use of a deep-learning based model to automatically segment artefacts from rNCCT maps. Our model circumvents time-invasive manual removal of artefacts from the rNCCT map and thereby simplifies segmentation of the ischemic core on baseline NCCT. Validation with external datasets is necessary before use in routine stroke evaluation.