Acute ischemic stroke caused by large vessel occlusion (LVO) carries high morbidity and mortality, but is readily treatable if diagnosed early. Expeditious diagnosis and triage of LVO patients to hospitals equipped to perform emergency revascularization treatment maximizes the patient’s chance of good clinical outcome. Current clinical LVO screening tools either lack precision or require specialized personnel to perform comprehensive clinical assessment. The objective of the present study is to develop a novel, rapid and automated computer algorithm capable of detecting and predicting signs and likelihood of LVO.
A preliminary model developed with machine learning techniques utilized clinical and imaging data of 300 patients provided by the Hospital Authority. Building on the pilot results, we further fine-tuned the computer algorithm with more patient data and imaging parameters derived from contrast CT-angiograms and CT brain scans interpreted by a team of Neuroradiologists to determine the ‘ground truth’ (i.e., the presence of absence of LVO in individual patients). Deep learning is performed with the information generated to establish a computer algorithm to predict the likelihood of LVO, as well as to assess the feasibility of computer interpretation of CT and CT-angiogram for signs of LVO. Here we present the developmental process and findings of this multi-disciplinary collaborative project, and discuss the potential to apply this innovation in clinical practice.

Background
There has been a lot of hype around the application of artificial intelligence (AI) in medical imaging. The number of publications on this subject has exploded lately. However, the vast majority of published studies were about development, validation and postulation of benefit of AI systems. Examples of actual usage of such systems are still scanty. We here share our early experience of actual usage of an AI system to triage non-urgent chest x rays.

Aim
The chest x-ray is the most commonly performed radiology investigation. Its sheer volume precludes reporting of all of them by radiologist in HA. And for exams reported, report turnaround time has much room for improvement. We aim to apply AI to triage non-urgent CXRs so that those exams requiring early attention could be reported earlier. 

Methodology
We used convolutional neural network based algorithms and leveraged a public dataset of CXRs to kickstart creation of a prediction model for CXRs. This model was applied to classify non-urgent CXR exams requested by general out-patient clinics into high or low risk such that priority reporting by radiologists can be done for high risk exams. We have performed retrospective virtual simulation of this AI-enabled priority reporting workflow followed by prospective pilot of such workflow in Hong Kong West Cluster.

Results
Both retrospective and prospective results showed that AI-enabled priority reporting workflow resulted in dramatic reduction in average report turnaround time for CXR exams that required early attention. This was achieved without additional radiologist manpower.

Conclusion
Employing AI to prioritize non-urgent CXRs translated into shorter report turnaround time for exams that required early attention.
 
Authors:
VSH Chan1, BXH Fang1, ITH Chen2, ESL Yau2, VKL Lau2, EYW Yeung2, PTM Lui2, JCH Wan2, AYH Tong2, MCY Cheng2, JCM Poon2, WWT Chan2, NT Cheung2, YC Wong3, DML Tse1, WWM Lam1, TPW Lam1

1 Department of Radiology, Queen Mary Hospital, Hong Kong
2 Information Technology and Health Informatics Division, Hospital Authority Head Office
3 Department of Radiology and Nuclear Medicine, Tuen Mun Hospital