ARRS 2022 Abstracts

RETURN TO ABSTRACT LISTING


1076. Artificial Intelligence (AI) Triage of Pneumothorax on Chest Radiographs: From Clinical Implementation to Turnaround Time Improvement
Authors * Denotes Presenting Author
  1. Neal Shah; Case Western Reserve University/University Hospitals
  2. Colin Marshall *; Case Western Reserve University/University Hospitals
  3. Gregory O'Connor; Case Western Reserve University/University Hospitals
  4. Jared Durieux; Case Western Reserve University/University Hospitals
  5. Kaustav Bera; Case Western Reserve University/University Hospitals
  6. Amit Gupta; Case Western Reserve University/University Hospitals
Objective:
The purpose of this study is to gauge the accuracy of an FDA-approved artificial intelligence (AI) tool for pneumothorax detection on chest radiographs in the real clinical environment and to analyze the potential improvement in radiologist workflow by means of assessing time to diagnosis following implementation of this tool on a routine clinical Picture Archive and Communication system (PACS).

Materials and Methods:
This retrospective study was approved by local institutional review board (IRB). Seamless integration of the AI tool into clinical PACS was already achieved before the commencement of this study. A cohort of 2016 patients with 7536 chest radiographs was identified from 09/2020 - 03/2021. The patients with AI alerts (TotalAI) were divided into two groups: flagged positive for pneumothorax (Ptx pos AI) and negative for pneumothorax (Ptx neg AI). A control group was identified, composed of patients scanned without the use of AI (No AI). The final radiologist report was taken as ground truth for diagnosis. The pneumothorax size, priority status as per the clinical order, and time to diagnosis (time of first dictation - time of study appearance on PACS) were also obtained.

Results:
Out of 6626 radiographs screened by the AI tool, 1212 (18.3%) were flagged positive for pneumothorax. Time to diagnosis for Ptx posAI group was substantially lower than the No AI group, which was statistically significant (133.3 vs 367 minutes, p < 0.01). There was also a difference between time to diagnosis for the Ptx posAI versus Ptx neg AI groups (133.3 vs 342.2 minutes, p< 0.01). Furthermore, there was decreased time to diagnosis in the Ptx posAI group with STAT priority versus routine/non STAT studies that were flagged by AI (74.7 vs 321.4 minutes, p< 0.001). There was no difference in time to diagnosis between studies flagged as priority 1 STAT versus priority 2 STAT (p = 0.74). The sensitivity and specificity of AI for clinically significant pneumothoraxes (moderate or large) were 85.1% and 87.7%, respectively, and for any size pneumothorax (small, moderate, or large) sensitivity and specificity values were 73.7% and 93.6%, respectively.

Conclusion:
Implementation of an AI tool into the clinical environment is possible and results in meaningful improvement in radiologists’ turnaround time, without compromising the diagnostic accuracy. Additionally, pairing AI with existing triage system of STAT/routine ordering can further improve the time to diagnosis. This study paves a way for further expanded application of such AI algorithms in clinical practice and proves that AI-powered triage can quicken identification and reporting of pneumothoraxes. Thus, adding AI tools can increase efficiency in radiology without negatively impacting accuracy.