Introduction
The global burden of intracerebral hemorrhage (ICH) is estimated at about 5 million events annually and has a high morbidity and mortality [
1]. Minimally invasive surgery (MIS) for hematoma evacuation aims to reduce ICH volume and perihematomal edema. Current guidelines recommend MIS in patients with supratentorial ICH > 20–30 ml volume and Glasgow coma scale of 5–12 [
2]. The MISTIE III trial is the largest study on MIS to date and identifies a reduction of the clot size to 15 ml or less and a correct drain position within the ICH as the aim for the procedure [
3]. Therefore, estimation of ICH volume is of therapeutic relevance.
In general, the ABC/2 method is widely established to measure volumes in CT scans. It was validated in different clinical settings e.g. neoplasia [
4‐
6]. Nevertheless, other investigations identified significant deviation from manual planimetric methods, especially in irregularly shaped objects [
7,
8]. Planimetric volumetry of ICH is a time intensive task averaging roughly 3.4 min per patient [
9].
Advancements in supervised machine learning using convolutional neural networks (CNN) for automated segmentation have demonstrated high accuracy in detecting, classifying, and segmenting ICH. However, these studies required large datasets of 300, 600 and 1732 scans respectively [
9‐
11]. In contrast, hierarchical, patch-based CNN architectures, trained on smaller datasets, enable segmentation in large 3D images, exhibiting encouraging results in complex segmentation tasks [
12‐
14].
In this study, our objective was to develop a machine learning algorithm for the segmentation of ICH and of the drain and volumetry of ICH subsequent to minimally invasive surgery of acute supratentorial ICH.
Materials and methods
Approval of the institutional review board was obtained and the requirement for informed consent was waived. We selected patients suffering from supratentorial ICH that were treated with MIS from a retrospective database. Inclusion criteria were age ≥ 18 years and available CT-imaging. No exclusions were made based on scanner model, settings, voxel size or presence of artefacts.
Ground truth (GT) annotation and development of the CNN were carried out using a local instance of the Nora imaging platform (
https://www.nora-imaging.com). Image calculations were done using MATLAB (MATLAB R2021a, The MathWorks). Statistical evaluation of the results and plotting were done using R software version 4.2.0 [
15].
Imaging datasets
Thirty nine scans from 29 patients examined between years 2011 and 2018 were randomly selected from our database. To avoid data leakage, we partitioned our data on the patient level, thus ensuring that repeat examinations of all patients were assigned to the same group. We randomly divided the data into training (n = 21 patients / 29 scans), validation (n = 3 patients / 4 scans) and testing (n = 5 patients / 6 scans). To avoid the effect of random patient selection on the results, we added not yet included consecutive patients examined between 2010—2012 to the testing dataset for a total of 59 scans belonging to 44 patients. The mean age (SD) was 70 (± 13.56) years and there were 36 male patients (52.9%).
Ground truth
Non-overlapping segmentation masks of the ICH and the intracranial part of the drain were manually delineated by a neuroradiologist, with three years of experience (A.E.). Overlapping voxels were subsequently identified and classified to their corresponding mask by applying a threshold operation with voxels ≥ 100 HU assigned to the drain mask.
CNN segmentation of ICH and drain
No preprocessing was applied to the CT data. The development of the CNN model relied on the Patchwork CNN Toolbox [
12]. Here, the input for the CNN was the CT image in HU units. Instead of normalizing/cropping the image, an initial channel splitting layer was used. This channel splitting layer separates the input value range into 11 feature channels that are sensitive to a particular HU range. This method was inspired by the windowing approach that a radiologist uses when reading images by dividing the entire HU area into detachable image parts, e.g., CT windows for soft tissue or bone. The ranges are initialized with the following centers [-1000, -500, -100, -50, -25, 0, 25, 50, 100, 500, 1000], and further refined during training. Three hierarchical scales (patches) were used. The finest scale was reformatted to 1-mm isotropic voxels.
To determine the best model parameters we initially tested six different combinations on 10
6 image patches, experimenting with two different versions of three model parameters.
1.
Feature dimensions in each scale: [8, 16, 16, 32, 64] or [8, 16, 16, 32, 64, 64]
2.
Loss function: categorical or binary cross-entropy.
3.
Augmentation at each level of the network: rotation angle 0.2, right-left flipping and zooming 10–20% or rotation angle 0.4, flipping in all dimensions, zooming 10–20% and random uniform scaling of the voxel values in each scale.
We employed the Deepmind library (
https://github.com/deepmind/surface-distance) to measure overlap and spatial distance metrics.
1.
Dice similarity coefficient (DSC) which measures the overlap of two sets of points
2.
Surface DSC, which measures the overlap of the surfaces of two sets of points at a specific tolerance (1 mm). The surface DSC is better suited than DSC for assessing performance in 3D segmentation tasks [
16].
3.
Surface overlap measures the average overlap at a specific tolerance (1 mm) returning two values. The average overlap from the GT surface to the predicted surface and vice versa.
4.
Hausdorff distance measures the distance between two sets of points. To alleviate its sensitivity to outliers, both the Hausdorff100 and Hausdorff95 (top 95% of the distances are taken into account) were evaluated.
5.
Average surface distance, which measures the distances between the surfaces of two sets of points at a specific tolerance (1 mm) and thus returning two values. The average distance from the GT surface to the predicted surface and vice versa [
16,
17].
The top-performing model on the validation dataset was trained using 1.2 × 107 patches. The model output is a 4D NIfTI object with two 3D 1-mm isotropic NIfTI volumes, indicating the probability of each voxel belonging to ICH/drain or to the background. Binary masks were produced using a threshold to optimize performance measures of the CNN. The volume of ICH was calculated by summing the 1-mm isotropic voxels of the ICH mask.
Comparison with no-new-U-Net (nnU-Net)
Isensee et al. [
18] published a self-adapting semantic segmentation method that was tested on a wide variety of medical imaging datasets with good results, as well as achieving top placements in multiple segmentation challenges. Using the same data partitions, we trained and tested an nnU-Net model using our datasets.
Statistical tests
To assess the agreement between predicted and GT ICH volume we calculated the intraclass correlation coefficient (ICC) [
19]. We also generated concordance plots and Bland–Altman plots [
20] to visualize the agreement between the two measurements.
Discussion
We developed a CNN machine learning model to segment ICH and drains in cases treated with minimally invasive surgery. Our model accurately segmented the ICH and drain with DSC scores of 0.86 and 0.91 respectively. Additionally, automated ICH volumetry yielded high agreement with ground truth (ICC = 0.94 [95% CI: 0.91, 0.97]), overestimating the ICH volume by 1.33 mL. We developed our model with relatively small training and validation datasets (n = 33) of heterogeneous data, originating from various scanners, a wide range of voxel sizes, and anisotropy, which enhances the model’s generalizability. Moreover, we did not employ image preprocessing, minimizing processing power demands, and making the model independent of preprocessing algorithm results such as skull stripping or cropping.
As ICH demonstrates excellent contrast in CT, it has been utilized for automated diagnosis with different levels of success. Most of the previous research focused on detecting ICH and reported accuracy measures reaching 0.98 for area under the receiver operating characteristic curve [
21]. In our study, all cases suffered an ICH and all were successfully detected. However, this was not the purpose of our study and we acknowledge a conceivable selection bias in our cohort. As we only included patients selected for MIS, the ICH volume in our cohort may have been skewed towards larger volume (41.5 mL). The average ICH volume in our cohort, while comparable to that reported in the MISTIE III Trial (47.4 mL) [
10], was higher compared to studies focusing on ICH segmentation. For instance, Ironside et.al. reported an average volume of 25.7 mL [
9].
In our study, we applied a patch based CNN toolbox [
12], which allows model development by creating training patches in the magnitude of millions from a small number of scans (n = 33 in our study). Testing on a larger test dataset (n = 59) resulted in a sufficient model performance (DSC = 0.86). The results are comparable with other published works, which reached a DSC of 0.92 training on several hundreds of scans [
11,
22‐
27]. Applying the no-new-Unet segmentation toolbox [
18] to our dataset yielded a good result (DSC = 0.87) that is comparable to our model. We assume that the sufficient results in spite of using a small dataset, may be attributed to the excellent image contrast in this segmentation task.
Volume reduction of ICH on follow up examinations is an aim of the treatment according to the MISTI III study and has been linked to 12-month mortality [
3,
28]. Accurate automated volumetry could address the shortcomings of the ABC/2 approach. However, to accomplish this, an accurate segmentation of the hemorrhage is required. We achieved a very high correlation between predicted and GT ICH volume (ICC = 0.94 [95% CI: 0.91, 0.97]), which lies within the range of other automated algorithms, where a comparable correlation coefficient reaching 0.98 was achieved [
11,
23,
24,
26,
27].
Furthermore, we achieved accurate segmentation of the drain (DSI = 0.91). A literature search yielded no other articles that attempt to segment drains following MIS. Precise segmentation of the ICH and drain might potentially simplify the evaluation of the drain position following MIS, which is crucial for treatment success.
Conclusions
Using a relatively small dataset, originating from different CT-scanners and with heterogeneous voxel dimensions, we applied a patch-based CNN framework and successfully developed a machine learning model, which accurately segments the intracerebral hemorrhage (ICH) and the drains. This provides automated and accurate volumetry of the bleeding in acute ICH treated with minimally invasive surgery.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit
http://creativecommons.org/licenses/by/4.0/.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.