Keywords:
Artificial Intelligence, Neuroradiology brain, CT, CT-Angiography, CAD, Computer Applications-Detection, diagnosis, Haemorrhage, Ischaemia / Infarction
Authors:
A.-A. El-Ahmadi, G. Brun, A. Ayobi, S. Quenet, Y. Chaibi, A. Reyre, A. Jacquier, N. Girard
DOI:
10.26044/ecr2024/C-13784
Methods and materials
All patients who underwent NCCT and/or CTA scans for stroke suspicion between March 2019 and July 2020 at two hospitals (La Timone and Hopital-Nord, Marseille, France) were retrospectively and consecutively collected. In order to establish the ground truth (GT), two board-certified neuroradiologists analyzed the scans and defined by consensus the presence of ICHs (and their subtypes: Intraparenchymal, Intraventricular and/or Extra-axial) and/or LVOs (and their locations in the Internal Carotid and/or Middle Cerebral Arteries segments: ICA, MCA-M1 and/or MCA-M2). In addition, if the case presented a LVO, the experts established the ASPECT Score.
Simultaneously, the data was processed by CINA-HEAD (Avicenna.AI, La Ciotat, France), which was previously integrated to flag suspected cases with ICH and/or LVO and to calculate the ASPECT Score. The algorithm results for ICH and LVO were compared to the expert consensus and the accuracy (overall agreement) was computed on a per-case basis. Similarly, for the ASPECTS analysis, a region-based accuracy was calculated. Moreover, a dichotomized ASPECTS analysis with a cutoff of ASPECTS ≥ 6 (the endovascular selection cut-off point) was performed by evaluating the percentage of cases with ASPECTS ≥ 6 according to the GT and identified as ASPECTS ≥ 6 by the algorithm.