At the emergency radiology department of the Universitair Ziekenhuis Brussel (UZ Brussel), an artificial intelligence (AI) tool supports on-call residents in detecting intracranial haemorrhage (ICH).
However, there is more to AI than just spotting hyperintensities on non-contrast head CT scans. An automated triage system integrated into the radiology workflow prioritises critical cases and ensures that those patients are diagnosed first, based on the patient’s severity and urgency.
In the future, AI will support the selection of imaging modalities, diagnoses, and even communication with colleagues and patients. However, this necessitates acknowledgment of the inherent variability and a good understanding of the use those applications in a clinical setting.
Based on my own experience as a radiologist in training at the UZ Brussel, I have become a strong proponent of hands-on training to understand the relevance and promising potential of AI in a clinical setting, during residency. I still remember one of my first nights on call when I received multiple head CTs in a short period of time. During the preliminary review, the AI tool already detected a subtle hyperdensity. After discussing it with a senior radiologist, an MRI confirmed a small haemorrhage in the region marked by the AI tool, that may otherwise have been overlooked.
Encouraged by this event, my team and I began to retrospectively collect a dataset, containing 500 consecutive head CT exams and their clinical reports, from September 1, 2019, to October 1, 2019. All CT exams were automatically pseudo-anonymised and transferred for ICH detection by an AI tool.
We registered the number of studies that were successfully processed by AI and calculated the diagnostic performance by the positive and negative predictive values, sensitivity, and specificity. The original clinical radiology report and consensus by independent supervision were considered the gold standard.
AI could automatically evaluate three-quarters of all head CT exams. With high specificity and negative predictive value, the AI tool shows the potential to rule out ICH, helping radiologists select which scans need to be viewed less urgently. The positive predictive value remains moderate.
To conclude, this is an opportunity to leverage the power of AI to shape the radiologists of the future. However, my research was not a one-person show; the success of implementation and operation was a collective effort from the whole team: the open mind of the Head of Radiology, Prof. Dr. Johan de Mey, the support and expertise of senior emergency radiologist Dr. Koenraad Nieboer, who put everything in motion, and the scientific approach and ever-critical eye of Prof. Nico Buls, medical physics expert.
Dr. Nina Watté is a radiologist in training at the UZ Brussel and a board member of the Young Radiologist Section (YRS) of the Belgian Society of Radiology.
Research Presentation Session
RPS 1117 Abdomen and brain
The performance of artificial intelligence in the detection of intracranial haemorrhage on head computed tomography with clinical workflow integration
N. Watté, K. H. Nieboer, N. Buls, J. de Mey; Brussels/BE
Read the full abstract in the ECR 2020 Book of Abstracts
Watté N, et al. (2020) The performance of artificial intelligence in the detection of intracranial haemorrhage on head computed tomography with clinical workflow integration. Abstract RPS 1117-3 in: ECR 2020 Book of Abstracts. Insights Imaging 11, 34 (2020). DOI 10.1186/s13244-020-00851-0