ERS3035. AI-Based Software for Automated ASPECTS Assessment Improves Physicians’ Performance
Authors * Denotes Presenting Author
  1. Mustafa Khan *; UC Irvine Medical Center
  2. Angela Ayobi; Avicenna
  3. Yasmina Chaibi; Avicenna
  4. Sarah Quenet; Avicenna
  5. Christopher Philippi; Tufts University Medical Center
  6. Dan Chow; UC Irvine Medical Center
  7. Peter Chang; UC Irvine Medical Center
The primary objective aims to evaluate physicians’ ability to assess Alberta Stroke Program Early CT Score (ASPECTS) when assisted by an AI-based automated software designed to identify hypodense regions on head non-contrast CT (NCCT) exams, compared with the unassisted readings. Secondly, this study aims to evaluate the standalone performance of the device.

Materials and Methods:
139 retrospective, consecutive and multicenter baseline NCCT from patients with acute middle cerebral artery and/or internal carotid artery occlusion were collected. The ASPECTS ground truth (GT) was established by consensus reading between 3 expert neuroradiologists. 40 cases were randomly selected from the main dataset for a multi-reader-multi-case (MRMC) study, conducted by 3 additional readers (radiologists different from the ground truthers) who assessed each exam first without software assistance and, after a washout period, with the assistance of CINA-ASPECTS (Avicenna.AI, La Ciotat, France). The readers used a 6-point confidence scale for the assessments of each ASPECTS region. Improvement in reader’s performance was determined by the percentage of the individual ASPECT regions where agreement was achieved between the reader and the GT, for both the unassisted and assisted reads. Moreover, the difference in the Receiver Operating Characteristic (ROC) Area Under the Curve (AUC) following the Obuchowski-Rockette-DBM-MRMC methodology was calculated. Finally, the standalone region-based performance of the device was evaluated against the GT for the 139 patients.

Regarding the MRMC study, all readers increased their agreements when assisted by the software: 79.8% vs 73.8% for reader 1 (p<0.05), 78.5% vs 76.5% for reader 2 (p>0.05) and 77.3% vs 71.8% for reader 3 (p>0.05). The overall difference among all readers was statistically significant: 78.5% vs 74.0% (p<0.05). Similarly, the use of CINA-ASPECTS significantly increased the average ROC AUC of all readers to 0.824 compared to a baseline average ROC AUC of 0.776 without software assistance (p<0.05). For the standalone performance, comparison of CINA-ASPECTS with the GT yielded a region-based sensitivity of 76.6% [95%CI: 72.4%–81.1%] and a specificity of 88.7% [95%CI: 87.4%–89.9%].

CINA-ASPECTS can identify hypodense brain regions with high concordance to expert neuroradiologists. In addition, the conjunctive use of the software as a diagnostic aid significantly improves the accuracy of physicians’ interpretations in a realistic clinical workflow.