AI shows good accuracy detecting intracranial haemorrhage on CT scans

A deep learning–based artificial intelligence (AI) model developed by researchers from the Chinese University of Hong Kong exhibits good accuracy in detection of intracranial haemorrhage (ICH) on CT scans.

Head CT scans constitute the main imaging investigation for evaluation of trauma and stroke. They are composed of multiple cross-sectional images (ie, slices), which may be challenging to interpret. Typically, these scans are initially reviewed by frontline physicians prior to assessment by radiologists, which can lead to delays in diagnosis confirmation.

“In the accident and emergency department, AI can facilitate ICH detection in head CT scans when a radiologist is unavailable,” wrote the researchers. “We developed a model using a publicly available international dataset of >750,000 expert-labelled CT slices and validated its performance on CT scans from our institution to determine its potential for clinical application in Hong Kong.” [Hong Kong Med J 2023;29:112-120]

The researchers acquired 25,312 head CT scans from the Radiological Society of North America (RSNA) open dataset, which were split into 752,807 slices ≥5 mm thick, then randomly shuffled and annotated by 60 volunteer experts from the American Society of Neuroradiology. Each CT slice was labelled to indicate the presence and type of ICH, namely, intraparenchymal haemorrhage (IPH), subarachnoid haemorrhage (SAH), subdural haemorrhage (SDH), epidural haemorrhage (EDH), and intraventricular haemorrhage (IVH). Approximately 14 percent of slices in the RSNA dataset were positive for ICH. [https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/data]

An AI model, which determines ICH probability for each CT scan and nominates five potential ICH-positive CT slices for review, was developed and validated using retrospective data from 1,372 noncontrast head CT scans (ICH-positive, 6.1 percent) collected at Prince of Wales Hospital in Hong Kong. The model achieved an area under the curve (AUC) of 0.842 (95 percent confidence interval [CI], 0.791–0.894; p<0.001) for scan-based detection of ICH. A prespecified probability threshold of ≥50 percent for the presence of ICH yielded 78.6 percent accuracy, 73 percent sensitivity, 79 percent specificity, 18.6 percent positive predictive value, and 97.8 percent negative predictive value. “Our model exhibited good accuracy in CT scan–based detection of ICH, considering the low prevalence of ICH in Hong Kong,” commented the researchers.

There were 62 true-positive scans and 22 false-negative scans, which were reduced to six false-negative scans by manual review of model-nominated CT slices. “With respect to true positives, most ICH-positive scans were detected; most of these scans had large areas of ICH, which presumably could be easily identified by nonradiologists. However, in six cases, the model correctly nominated CT slices with small areas of ICH,” highlighted the researchers.

“Considering the 6 percent prevalence of ICH in our institution, and using a prespecified probability threshold of ≥50 percent, the model detected 74 percent of ICH-positive scans, which improved to 93 percent via manual review of model-nominated images,” they reported.

The following AUC and corresponding sensitivity/specificity values were obtained for individual types of ICH at ≥50 percent probability threshold: IPH, 0.930 (95 percent CI, 0.892–0.968) and 4 percent/100 percent; SAH, 0.766 (95 percent CI, 0.684–0.849) and 12 percent/96 percent; SDH or EDH, 0.865 (95 percent CI, 0.783–0.947) and 75 percent/90 percent; IVH, 0.935 (95 percent CI, 0.852–1.000) and 85 percent/93 percent. “Our results support further development of the model to improve its accuracy and incorporate a mechanism that can facilitate visual confirmation of ICH location,” noted the researchers.