Deep Learning for Automated Measurement of Hemorrhage and Perihematomal Edema in Supratentorial Intracerebral Hemorrhage

Abstract

Background and Purpose- Volumes of hemorrhage and perihematomal edema (PHE) are well-established biomarkers of primary and secondary injury, respectively, in spontaneous intracerebral hemorrhage. An automated imaging pipeline capable of accurately and rapidly quantifying these biomarkers would facilitate large cohort studies evaluating underlying mechanisms of injury. Methods- Regions of hemorrhage and PHE were manually delineated on computed tomography scans of patients enrolled in 2 intracerebral hemorrhage studies. Manual ground-truth masks from the first cohort were used to train a fully convolutional neural network to segment images into hemorrhage and PHE. The primary outcome was automated-versus-human concordance in hemorrhage and PHE volumes. The secondary outcome was voxel-by-voxel overlap of segmentations, quantified by the Dice similarity coefficient (DSC). Algorithm performance was validated on 84 scans from the second study. Results- Two hundred twenty-four scans from 124 patients with supratentorial intracerebral hemorrhage were used for algorithm derivation. Median volumes were 18 mL (interquartile range, 8-43) for hemorrhage and 12 mL (interquartile range, 5-30) for PHE. Concordance was excellent (0.96) for automated quantification of hemorrhage and good (0.81) for PHE, with DSC of 0.90 (interquartile range, 0.85-0.93) and 0.54 (0.39-0.65), respectively. External validation confirmed algorithm accuracy for hemorrhage (concordance 0.98, DSC 0.90) and PHE (concordance 0.90, DSC 0.55). This was comparable with the consistency observed between 2 human raters (DSC 0.90 for hemorrhage, 0.57 for PHE). Conclusions- We have developed a deep learning-based imaging algorithm capable of accurately measuring hemorrhage and PHE volumes. Rapid and consistent automated biomarker quantification may accelerate powerful and precise studies of disease biology in large cohorts of intracerebral hemorrhage patients.

Keywords: biology; biomarkers; brain edema; cerebral hemorrhage; deep learning.

Figures

Figure 1:

Concordance of manual with automated hemorrhage (top) and perihematomal edema (bottom) volumes. Results from the cross-validation (Yale cohort) are shown in panels A and C while results from the external (ERICH) cohort are shown in panels B and D. Line of identity is also plotted.

Figure 2:

Bland-Altman plots of automated measurement of hemorrhage volume (top) and perihematomal edema volume (bottom) compared with manual ground-truth for the entire derivation cohort (panels A and C) and for the validation (ERICH) cohort (panels B and D). Dotted lines represents limits of agreement (1.96 times standard deviation).

Figure 3:

Dice similarity coefficients for segmentation of hemorrhage and perihematomal edema comparing manual to automated algorithm (red for cross-validation, green for external validation cohorts) and repeat testing by the same rater (i.e. intra-rater reliability, blue) and different rater (inter-rater, purple)