Authors (including presenting author) :
VSH Chan(1), BXH Fang(1), ITH Cheng(2), ESL Yau(2), VKL Lau(2), EYW Yeung(2), PTM Lui(2), JCH Wan(2), AYH Tong(2), MCY Cheng(2), JCM Poon(2), WWT Chan(2), NT Cheung(2), YC Wong(3), DML Tse(1), TPW Lam(1), WWM Lam(1)
Affiliation :
(1) Department of Radiology, Queen Mary Hospital, (2) Information Technology and Health Informatics Division, Hospital Authority Head Office, (3) Department of Radiology and Nuclear Medicine, Tuen Mun Hospital
Introduction :
Ever-increasing demand for radiology services in the public sector is met with a shortage of radiologists, and prolonged exam report turnaround time is one of the resultant manifestations.
Objectives :
We aim to demonstrate that using artificial intelligence (AI) to triage chest radiograph exams in the radiology department could translate into shortened report turnaround time for exams that warrant early clinical attention.
Methodology :
Using a publicly available dataset of chest radiographs — NIH dataset of 112120 chest radiographs, we trained a convolutional neural network based system to recognize 14 different disease categories on the frontal chest radiograph, achieving an average AUC under the ROC curve of 0.84. With this system, frontal chest radiographs could be classified into those that require early attention (high risk) and those that do not (low risk).
We then retrospectively retrieved 1738 consecutive non-urgent frontal chest radiographs requested by general out-patient clinics within Hong Kong West Cluster in October 2017 to March 2018. These exams were classified according to the radiology report into three categories – Category 0: Early attention not required (N=1362); Category 1: Early attention may be needed (N=228); Category 2: Early attention definitely warranted (N=138). Actual report turnaround time for each exam was charted.
Finally, we applied the algorithm to categorize the above exams into high or low risk and simulated a new reporting workflow whereby radiographs flagged by the system were given priority reporting. Exams in 2017 were used to optimize system parameters and testing was performed on the exams in 2018. A new set of report turnaround times was obtained.
Result & Outcome :
Results: For the exams in 2018, mean actual report turnaround times for category 0, 1 and 2 exams were 28.4, 34.3 and 34.1 days respectively. These numbers changed to 32.2, 22.9 and 12.7 days respectively after simulated application of priority reporting workflow. There was no change in workload of the radiologist as constrained in the simulation. The system has been deployed for prospective clinical use in December 2018 and latest prospective data will be also presented. Conclusion: Positive clinical impact in the form of shortened report turnaround time for chest radiographs that deserve early attention can be effected by the aid of an AI triage system.
Co-location with:
