gms | German Medical Science

Objective: The role of surgery in the treatment of ICH still remains elusive despite recent large randomized trials. Distinct factors predisposing patients with ICH to the risk of secondary neurological deterioration are still missing. Intracranial compartments (i.e. CSF, parenchyma, clot) are discussed to be of predictive value, exact volumetric analysis has not yet been performed however. A fully automatic segmentation algorithm in brain computed tomography (CT) is introduced here, allowing for volumetric analysis while leaving out the time consuming procedure of manual segmentation.

Method: A fully automatic segmentation algorithm implementing first- and second order statistics, texture features and automated threshold features was trained with a random forest methodology. For training, 13 randomly assorted brain CTs with spontaneous ICH were manually labeled for regions of ICH and intracranial CSF by two independent raters. In 13 additional subject-CTs, results from automatic segmentation were compared with manual segmentation and ICH volume acquired by the A*B*C/2 method.

Results: Automatic segmentation of subject-CTs yielded mean ICH volumes of 40,39 ± 28,53ml and 94,36 ± 71,47ml for CSF volume. In manual segmentation, mean ICH volume was 46,31 ± 30,76ml, mean CSF volume was 92,27 ± 57,0ml. Results from automatic segmentation showed strong systematic correlation with manual segmentation for ICH (r=0,98, p<0,0001) and CSF volume (r=0,95, p<0,0001). Mean ICH volume by A*B*C/2 method was 59,70 ± 44,32ml. In ANOVA, volumes by A*B*C/2 were significantly larger compared to either manual or automatic segmentations (p<0,0001). In contrast, ICH volumes of manual compared to automatic segmentations showed no significant difference (ANOVA for A*B*C/2 vs. automatic vs. manual segmentation followed by Benferroni's Multiple Comparison).

Conclusions: A fully automatic segmentation algorithm for CT-based volumetry is presented here serving as a prerequisite for subsequent analysis of intracranial volume proportions in the context of spontaneous ICH. In native brain CT, the segmentation algorithm could reliably predict volumes of ICH and CSF and indicates that ICH volumes are likely to be overestimated by the A*B*C/2 method. Automatic volume estimations were as precise as manual measurements, however acquired in a fraction of time. Quick and accurate information about the extent of ICH and the amount of spare CSF could help to identify patients at risk for secondary neurological deterioration.