gms | German Medical Science

71. Jahrestagung der Deutschen Gesellschaft für Neurochirurgie (DGNC)
9. Joint Meeting mit der Japanischen Gesellschaft für Neurochirurgie

Deutsche Gesellschaft für Neurochirurgie (DGNC) e. V.

21.06. - 24.06.2020

Performance of an advanced deep learning model in fully automated detection and segmentation of intracranial aneurysms in patients with subarachnoid haemorrhage

Automatische Detektion und Segmentierung von intrakraniellen Aneurysmen in Patienten mit Subarachnoidalblutung durch ein Deep Learning Model

Meeting Abstract

  • presenting/speaker Lukas Goertz - Universitätsklinikum Köln, Köln, Deutschland
  • Jan Borggrefe - Universitätsklinikum Köln, Köln, Deutschland
  • Rahil Shahzad - Universitätsklinikum Köln, Köln, Deutschland
  • Roland H. Goldbrunner - Universitätsklinikum Köln, Köln, Deutschland
  • Lenhard Pennig - Universitätsklinikum Köln, Köln, Deutschland
  • Boris Krischek - Universitätsklinikum Köln, Köln, Deutschland

Deutsche Gesellschaft für Neurochirurgie. 71. Jahrestagung der Deutschen Gesellschaft für Neurochirurgie (DGNC), 9. Joint Meeting mit der Japanischen Gesellschaft für Neurochirurgie. sine loco [digital], 21.-24.06.2020. Düsseldorf: German Medical Science GMS Publishing House; 2020. DocV270

doi: 10.3205/20dgnc266, urn:nbn:de:0183-20dgnc2664

Published: June 26, 2020

© 2020 Goertz et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License. See license information at



Objective: Aneurysmal subarachnoid haemorrhage (aSAH) is a potentially life-threatening condition that requires early aneurysm embolization in order to prevent rebleeding. Initial diagnosis of the aneurysm is usually made by computed tomography angiography (CTA), however, misdiagnosis may occur, particularly in association with physician fatigue and lacking experience. Convolutional neural networks have proven great potential in performing diagnostic diagnostic and analyzing tasks as assistance for clinicians. The objective was to develop a deep learning model (DLM) that detects and segments intracranial aneurysms fully automatically on CTA in patients with aSAH.

Methods: All patients received digital subtraction angiography to confirm the aneurysms. Sixty-eight patients with 79 aneurysms (2016-2017) served as the training cohort. Three DLMs with different architecture were trained using 5-fold cross-validation and their respective outputs were merged via an ensemble strategy (DLM-Ens). The DLM-Ens was evaluated on two independent validation cohorts (2010-2015): Cohort 1: 101 patients, 112 aneurysms, Cohort 2: 84 patients, 104 aneurysms. Independent manual segmentations of the aneurysms was performed in a voxel-wise manner by an experienced radiologist and a neurosurgeon and served as the ground truth.

Results: In the combined validation cohort, the detection sensitivity of the DLM-Ens for intracranial aneurysms was 0.82 (Cohort 1: 0.84, Cohort 2: 0.81) and the average number of false-positives per scan was 0.81 (Cohort 1: 0.78, Cohort 2: 0.82). The overall median dice coefficient was 0.75 (Cohort 1: 0.75, Cohort 2: 0.73), representing the segmentation accuracy. Detection sensitivity was 0.90 for aneurysms > 50 mm3 and 0.96 for aneurysms > 100 mm3. Aneurysm location (anterior/posterior circulation, p=0.07) and Fisher grade (grade ≤ 3 vs. grade 4; p=0.33) had no impact on diagnostic accuracy. There was a strong correlation between automated aneurysm segmentation and the ground truth, with a Pearson’s correlation coefficient of 0.95 for volume and 0.78 for maximum aneurysm diameter.

Conclusion: The created DLM detects aneurysms with sufficient sensitivity and independent of aneurysm location and bleeding severity, hence, convolutional neuronal networks may provide a valuable adjunct for radiologists during emergency imaging of patients with SAH.