Purpose
To implement an algorithm which detects intracranial haemorrhages on NCCT head studies.
To increase algorithm utility by subtyping detected haemorrhages and locating them.
To increase algorithm explainability by highlighting image pixels which contribute most significantly to the model's final prediction.
To demonstrate validity by evaluating performance on an independent test dataset.
Methods and materials
BACKGROUND
Artificial intelligence and deep learning
Deep learning (DL) is based on artificial neural networks, which train themselves on large amounts of data to perform a task.
By automatically discovering image patterns, they can be applied to detect or classify features in images (“computer vision”).
Intracranial haemorrhage
Intracranial haemorrhage is a time-critical neurological emergency with 5 subtypes: intracerebral (ICH), intraventricular (IVH), subarachnoid (SAH), subdural (SDH) and epidural (EDH).
Nearly half of the mortality from ICH occurs in the first 24 hours.1
Non-contrast computed tomography (NCCT)...
Results
The model, trained on dataset A, was tested on independent data from dataset B.
Evaluation of performance (Figures 2-3)
Detection of any haemorrhage (all subtypes) performed most consistently across all metrics (accuracy = 0.90, specificity = 0.87, sensitivity (recall) = 0.94, precision = 0.84), compared to the detection of specific subtypes.
EDH subtyping had the poorest performance, with the lowest sensitivity (0.77), precision (0.48) and smallest area under the precision-recall curve (0.74). This may be attributed to its low representation within the training dataset, reducing...
Conclusion
The current study demonstrated the technical feasibility of our DL implementation for the automatic detection of intracranial haemorrhages on NCCT head studies, with its performance validated on an independent dataset. It also presented effective approaches which increased our algorithm’s explainability and potential clinical utility.
However, our model performed less well on subtype classification, with reduced performance on less common subtypes such as EDH.
Further work will revolve around:
improving the algorithm’s performance on subtype detection, and
evaluating the model on prospective data (from the target...
References
Elliott J, Smith M. The Acute Management of Intracerebral Hemorrhage: A Clinical Review. 2010. p. 1419-27.
Alobeidi F, Aviv RI. Emergency Imaging of Intracerebral Haemorrhage. Front Neurol Neurosci. 2015;37:13-26
Cordonnier C, Demchuk A, Ziai W, Anderson CS. Intracerebral haemorrhage:. 2018;392(1015current approaches to acute management. The Lancet4):1257-68.