Background and Purpose: Reliable recognition of large vessel occlusion (LVO) on noncontrast computed tomography (NCCT) may accelerate identification of endovascular treatment candidates. We aim to validate a machine learning algorithm (MethinksLVO) to identify LVO on NCCT. Methods: Patients with suspected acute stroke who underwent NCCT and computed tomography angiography (CTA) were included. Software detection of LVO (MethinksLVO) on NCCT was tested against the CTA readings of 2 experienced radiologists (NR-CTA). We used a deep learning algorithm to identify clot signs on NCCT. The software image output trained a binary classifier to determine LVO on NCCT. We studied software accuracy when adding National Institutes of Health Stroke Scale and time from onset to the model (MethinksLVO+). Results: From 1453 patients, 823 (57%) had LVO by NR-CTA. The area under the curve for the identification of LVO with MethinksLVO was 0.87 (sensitivity: 83%, specificity: 71%, positive predictive value: 79%, negative predictive value: 76%) and improved to 0.91 with MethinksLVO+ (sensitivity: 83%, specificity: 85%, positive predictive value: 88%, negative predictive value: 79%). Conclusions: In patients with suspected acute stroke, MethinksLVO software can rapidly and reliably predict LVO. MethinksLVO could reduce the need to perform CTA, generate alarms, and increase the efficiency of patient transfers in stroke networks.

Artificial Intelligence (AI) can assist in vessel occlusion (VO) identification in acute stroke patients. We aim to investigate the impact of using an AI-based software for automated VO detection on non-contrast CT (AI-VO) as compared to CT-Aangiograpphy (CTA). Methods: From April to October 2020 all patients admitted with a suspected acute ischemic stroke underwent urgent non-contrast CT / CTA / CTP and were treated accordingly. Hypoperfusion areas defined as Tmax>6s on CTP (RAPID software), congruent with the clinical symptoms and a vascular territory, were considered VO (CTP-VO: ground truth). In addition, two experienced neuroradiologists blinded to CTP but not to clinical symptoms retrospectively evaluated CT and CTA to identify intracranial VO (CTA-VO). AI-VO was automatically determined by an AI-based software (Methinks). Results: Of the 338 patients included, 157 (46.5%) showed a CTP-VO (median Tmax>6s: 73[29-127]ml). Overall sensitivity to detect CTP-VO was 50.3% for CTA-VO and 66.9% for AI-VO; specificity was 97.8% for CTA-VO and 86.2% for AI-VO. EVT was performed in 103 patients (EVT-VO: 65.6% of CTP-VO; Tmax>6s: 102[63-160]ml); sensitivity to detect EVT-VO was 69% for CTA-VO and 79.6% for AI-VO; specificity was 95.3% for CTA-VO and 79.6% for AI-VO. The probability to detect a CTP-VO was higher with AI than with CTA for distal occlusions (figure). Accordingly, AI-VO sensitivity was higher than CTA-VO for angiographically confirmed M2/M3-MCA occlusions (80.7% vs 34.6%; p=0.002) but not for M1-MCA/ICA occlusions (82.1% Vs 88.1%;p=0.467). Conclusion: AI-assisted vessel occlusion identification on non-contrast CT may be a useful tool in acute stroke evaluation, especially for distal VO identification, potentially increasing endovascular treatment in these cases.

Purpose: To validate a Machine Learning algorithm able to identify LVO on NCCT. Methods: Patients with suspected acute stroke who underwent NCCT+CT Angiography (CTA) from two comprehensive stroke centers were included. Patients with intracranial haemorrhage were excluded. Two experienced radiologists identified the presence of LVO on CTA (NR-CTA) tagging the clot location and manually segmenting the clot. Acute ischemia and clot signs on NCCT were also depicted with assistance of the CTA clot location. With this information a deep learning system was used to create an algorithm (Deepstroke) to identify and locate the presence/absence of acute ischaemia and clot signs in NCCT. Deepstroke image output was used to train a binary classifier to determine LVO on NCCT. Cross-validation was performed in a stratified 5-fold of the data, including deep learning training. We also studied the effect on Deepstroke accuracy when adding the patients NIHSS and time from onset to the model (Deepstroke+). Results: The data cohort included 1354 patients, 724 (53%) with LVO by NR-CTA. The accuracy of Deepstroke to identify LVO had an AUC of 0.81 (sensitivity 0.85; specificity 0.49, PPV 0.66, NPV 0.74), and improved combined with NIHSS and time from symptom onset to AUC 0.88 (sensitivity 0.87, specificity 0.68, PPV 0.76, NPV 0.82). Deepstroke performed better on larger occlusions (Table). Among patients identified as LVO by Deepstroke+ only 19% showed no findings on NR-CTA. The agreement in LVO detection between NR-CTA and Deepstroke+ was 0.78 (Deepstroke was 0.68). Process time per patient was below 120s. Conclusions: In patients with suspected acute stroke, Deepstroke identified LVO in NCCT with a high correlation with radiologist readings of CTAs. Deepstroke could reduce the need to perform CTA, generate alarms and increase the efficiency of patients transfers in the acute management in stroke networks. Deepstroke accuracy will improve as more cases are added to the training set.