Deep learning-based automated segmentation of intracerebral haemorrhage, intraventricular haemorrhage and perihaematomal oedema on non-contrast CT

TL;DR

This paper presents a deep learning model for automatically segmenting brain hemorrhage and related conditions in CT scans, showing high accuracy for some regions but lower for others.

Contribution

A novel deep learning model for simultaneous segmentation of ICH, IVH, and PHO on non-contrast CT scans is developed and validated.

Findings

The model achieved a median Dice coefficient of 0.93 for intracerebral haemorrhage segmentation.

Segmentation of intraventricular haemorrhage had a median Dice coefficient of 0.75.

Perihaematomal oedema segmentation showed moderate performance with a Dice coefficient of 0.53.

Abstract

Precise volumetric evaluation of intracerebral haemorrhage (ICH), intraventricular haemorrhage (IVH) and perihaematomal oedema (PHO) is essential but manual segmentation is time-consuming and susceptible to variability. We aimed to develop and externally validate a deep learning model for simultaneous segmentation of ICH, IVH and PHO on non-contrast CT (NCCT) in patients with spontaneous ICH.

A 3D U-net model was trained with 5-fold cross-validation on baseline NCCTs from 301 patients included in 2 prospective multicentre studies. External validation was performed on 141 baseline NCCTs from another multicentre study. Model performance was evaluated against manual ground truth segmentations using the Dice similarity coefficient (DSC), intraclass correlation coefficients (ICC) and Bland–Altman analyses.

The model achieved a median DSC of 0.93 (IQR 0.91–0.94) for ICH, 0.75 (IQR 0.57–0.82) for IVH and 0.53 (IQR 0.34–0.65) for PHO. Volume correlations were excellent for ICH (mean absolute and consistency ICC both 0.98 [95% CI 0.98–0.99]) and IVH (absolute ICC 0.97 [95% CI 0.92–0.98]; consistency ICC 0.98 [95% CI 0.96–0.99]), and moderate for PHO (absolute ICC 0.60 [95% CI -0.08–0.85]; consistency ICC 0.82 [95% CI 0.76–0.87]). Bland–Altman analyses demonstrated a bias for ICH of −0.48 mL (LoA −8.21 to 7.26), for IVH of −1.68 mL (LoA −7.35 to 3.99) and for PHO of 13.91 mL (LoA −4.85 to 32.68).

The model enables accurate automated segmentation of ICH, while IVH and PHO segmentation remain more challenging. Automated segmentations may already serve as reliable pre-segmentations in research, but require visual assessment and correction, in particular for IVH and PHO.

Graphical Abstract

Figures