Deep Residual Dilated Convolutional Learning for Detection of Large Vessel Occlusion in Ischemic Stroke Patients

Rapid identification of large vessel occlusions (LVOs) is crucial when treating and managing patients with acute ischemic strokes (AIS). This urgency is due to the fact that LVOs are associated with high rates of post-stroke dependence and mortality. Furthermore, precise and prompt detection of such occlusions can assist in minimizing long-term functional and psychological ramifications. The conventional method for LVO detection requires manual examination of Computed Tomography Angiography (CTA) brain images. However, treatment delays frequently occur due to the unavailability of skilled radiologists specialized for these cases. In this study, we introduce a Three-Dimensional Deep Residual Dilated Convolutional Neural Network (DRDCNet-3D) that can harness the capabilities of dilated convolutions to improve spatial resolution. This method allows for rapid and precise detection of LVOs directly from three-dimensional CTA brain scans. Utilizing the recently published IACTA-EST dataset, our designed model has exhibited remarkable performance, boasting an AUC-ROC of 0.91 and an F1-score of 0.90. Such performance represents an average improvement of over 20% compared to existing state-of-the-art 3D medical image analysis models available in open-source platforms. This notable enhancement underscores the model's efficacy in diagnosing large vessel occlusions (LVOs). Our approach equips physicians with valuable insights for determining optimal treatment strategies in suspected AIS cases. Accordingly, timely interventions can be carried out, reducing the risk of permanent severe damage and benefiting patient outcomes.