Automated segmentation of haematoma and perihaematomal oedema in MRI of acute spontaneous intracerebral haemorrhage

Abstract

Background: Spontaneous intracerebral haemorrhage (SICH) is a common condition with high morbidity and mortality. Segmentation of haematoma and perihaematoma oedema on medical images provides quantitative outcome measures for clinical trials and may provide important markers of prognosis in people with SICH.

Methods: We take advantage of improved contrast seen on magnetic resonance (MR) images of patients with acute and early subacute SICH and introduce an automated algorithm for haematoma and oedema segmentation from these images. To our knowledge, there is no previously proposed segmentation technique for SICH that utilises MR images directly. The method is based on shape and intensity analysis for haematoma segmentation and voxel-wise dynamic thresholding of hyper-intensities for oedema segmentation.

Results: Using Dice scores to measure segmentation overlaps between labellings yielded by the proposed algorithm and five different expert raters on 18 patients, we observe that our technique achieves overlap scores that are very similar to those obtained by pairwise expert rater comparison. A further comparison between the proposed method and a state-of-the-art Deep Learning segmentation on a separate set of 32 manually annotated subjects confirms the proposed method can achieve comparable results with very mild computational burden and in a completely training-free and unsupervised way.

Conclusion: Our technique can be a computationally light and effective way to automatically delineate haematoma and oedema extent directly from MR images. Thus, with increasing use of MR images clinically after intracerebral haemorrhage this technique has the potential to inform clinical practice in the future.

Keywords: Brain MRI; Image segmentation; Spontaneous intracerebral haemorrhage; Stroke.

Fig. 1

Pre-processing workflow including registration and generation of the different brain masks used in the proposed method. Green boxes correspond to inputs and yellow boxes correspond to outputs.

Fig. 2

Late subacute phase MR sequences of three different SICH patients. (a)–(c): T2* GRE. (d)–(f): FLAIR. Note that on FLAIR imaging the haematoma can be completely iso-intense to grey matter, completely hypo-intense, or hypo-intense with an iso-intense centre. On the contrary, the haematomas tend to consistently appear hypo-intense in the T2* GRE images, albeit with a possible small iso-intense centre.

Fig. 3

Linear patterns for detecting weak connections. We set the centre voxel to false if its 26-connected neighbourhood forms the (a), (b) or (c) pattern. The 20 neighbours that are always false in all three patterns are not depicted for ease of visualisation.

Fig. 4

Haematoma segmentation workflow. Green boxes correspond to inputs and yellow boxes correspond to outputs.

Fig. 5

Visualisation of geodesic distance. (a) Simulated example of haematoma (grey) and hyper-intensity (white) segmentation. (b) Geodesic quasi-euclidean distance map from haematoma for each voxel belonging to hyper-intensity. Note that the distance is the shortest separation measured along the hyper-intense segmentation.

Fig. 6

Perihaematoma oedema segmentation workflow. Green boxes correspond to inputs and yellow boxes correspond to outputs.

Fig. 7

Box-and-violin plot of aggregated pairwise Dice overlaps between all 5 raters (Human vs Human) and aggregated Dice overlaps between the proposed algorithm and every expert rater (Human vs Proposed).

Fig. 8

Box-and-violin plot of Dice overlaps for the proposed automated segmentation with respect to each rater over all 18 patients.

Fig. 9

Visualisation of haematoma and perihaematoma oedema ($\lambda =15$) segmentation on Dataset A, overlaid on the FLAIR image. Worst case (left column), median case (middle column) and best case (right column), considering the average Dice score over all raters. Manual segmentations in rater order from top to bottom, with the bottom row being the proposed automated segmentation. Best viewed in colour.

Fig. 10

Worst case of haematoma segmentation on Dataset A. (a) FLAIR, axial. (b) FLAIR, coronal. (c) FLAIR, sagittal. Note that there is over-segmentation (yellow arrows) due to attenuated FLAIR intensities within oedema.

Fig. 11

Box-and-violin plot of Dice overlaps for DeepMedic vs the proposed automated segmentation using random half-partitions of datasets A and B for training and testing.

Fig. 12

Visual results showing two Dataset B cases where a similarly good performance between the proposed method and DeepMedic is observed for both haematoma (top) and oedema (bottom). Left column: Manual segmentation. Middle column: DeepMedic segmentation. Right column: Proposed method.

Fig. 13

Axial, coronal and sagittal views of the proposed registration-based (top row) and re-labelled MALP-EM (bottom row) label maps of case 5, overlaid on the T1 image. We observe that re-labelled MALP-EM label map has a larger extent (see arrows in registration-based map for comparison), which resulted in a “leakage” of the T2* GRE region of interest and subsequent erroneous localisation of haematoma.

Fig. 14

Segmentations overlaid on the T1 image for case 1 (top row) and case 2 (bottom row). Left column corresponds to the segmentation using the re-labelled MALP-EM map, middle column corresponds to the segmentation using the proposed registration-based map, and right column corresponds to the majority-vote label fusion amongst the five expert raters. Note that the re-labelled MALP-EM maps have non-ventricular labels for voxels corresponding to intraventricular haemorrhage, causing haematoma over-segmentation (arrows).

Fig. 15

Segmentations overlaid on the T1 image for case 4. Left column corresponds to the segmentation using the re-labelled MALP-EM map, middle column corresponds to the segmentation using the proposed registration-based map, and right column corresponds to the majority-vote label fusion amongst the five expert raters. Note that the proposed registration-based label map incorrectly labels part of the haematoma as being within the fourth ventricle, resulting in under-segmentation of haematoma (arrow).