Skip links

Clinical Study

Detection of intracranial haemorrhage on CT of the brain using a deep learning algorithm

D. Desbuquoit, S. Dekeyzer, J. Huyskens, S. Nicolay, E. De Smet, J. Van Goethem, L. Van den Hauwe, P. M. Parizel; Edegem/BE

Purpose: This prospective study aimed to evaluate the use of a commercially available deep learning algorithm for the detection of intracranial haemorrhage on non-contrast enhanced CT of the brain.
Methods and Materials: 500 non-contrast enhanced CT’s of the brain performed in June, July and August 2018 were independently analysed on the presence of pathological hyperdensities by a deep learning software package (Aidoc, Tel Aviv, Israel) and a 4th-year radiology resident. Their results were compared to a “gold standard analysis”, performed by a senior neuroradiologists with access to clinical information and, when available, previous or follow-up imaging studies.
Results: Pathological hyperdensities were present in 134/500 patients, the majority of which were haemorrhages (128/134; 95.5%). Pathological hyperdensities were correctly identified by Aidoc-software in 125/134 cases (93,3%), compared to 133/134 (99.3%) for the resident. Aidoc’s false-negative ratio was 9/134 (6.7%). When no pathological hyperdensities were present, the exam was rated negative by Aidoc-software in 345/366 cases (94.3%), compared to 362/366 (98.9%) for the resident. Aidoc’s false-positive ratio was 21/366 (5.7%).
Conclusion: The use of a deep learning algorithm for the detection of pathological intracranial hyperdensities helped to detect urgent cases more quickly. False positive results occur in a limited number of cases (5,7%) and are mainly due to beam hardening artefacts, hyperdense dural sinuses, or falcine or basal ganglia calcifications. False negatives were slightly more frequent (6,7%) and mainly seen in small or subtle haemorrhages.