Quantitative analysis of brain herniation from non-contrast CT images using deep learning

https://doi.org/10.1016/j.jneumeth.2020.109033Get rights and content

Highlights

  • New predictors responsible for brain herniation is explored.

  • Error is computed in pixels, area and volume.

  • Volumetric relation of brain shift is explored.

Abstract

Background

Brain herniation is one of the fatal outcomes of increased intracranial pressure (ICP). It is caused due to the presence of hematoma or tumor mass in the brain. Ideal midline (iML) divides the healthy brain into two (right and left) nearly equal hemispheres. In the presence of hematoma, the midline tends to shift from its original position to the contralateral side of the mass and thus develops a deformed midline (dML).

New method

In this study, a convolutional neural network (CNN) was used to predict the deformed left and right hemispheres. The proposed algorithm was validated with non-contrast computed tomography (NCCT) of (n = 45) subjects with two types of brain hemorrhages - epidural hemorrhage (EDH): (n = 5) and intra-parenchymal hemorrhage (IPH): (n = 40)).

Results

The method demonstrated excellent potential in automatically predicting MLS with the average errors of 1.29 mm by location, 66.4 mm2 by 2D area, and 253.73 mm3 by 3D volume. Estimated MLS could be well correlated with other clinical markers including hematoma volume - R2 = 0.86 (EDH); 0.48 (IPH) and a Radiologist-defined severity score (RSS) - R2 = 0.62 (EDH); 0.57 (IPH). RSS was found to be even better correlated (R2 = 0.98 (EDH); 0.70 (IPH)), hence better predictable by a joint correlation between hematoma volume, midline pixel- or voxel-shift, and minimum distance of (ideal or deformed) midline from the hematoma (boundary or centroid).

Conclusion

All these predictors were computed automatically, which highlighted the excellent clinical potential of the proposed automated method in midline shift (MLS) estimation and severity prediction in hematoma decision support systems.

Introduction

A healthy human brain is (nearly) symmetric with an imaginary line (passing between the lateral ventricles) – called midline – dividing the brain into symmetric left and right hemispheres. The presence of hematoma or tumor mass causes enlargement of ipsilateral brain tissue, which in turn causes the mid-brain and ideal midline (iML) to shift towards the contralateral side leading to deformed midline (dML). The presence of a mass within the periphery of the healthy brain tissue increases intracranial pressure (ICP) that leads to dML (see Fig. 1). According to the Monro-Kellie hypothesis (Mokri, 2001) - any space-occupying lesion (SOL) such as gliomas, hematoma, tumor, etc., causes mid-brain shift followed by herniation increases ICP, compresses the brainstem, and finally death. The volume of hematoma, amount of midline shift (MLS), etc., are the primary factors of brain compression. Higher MLS leads to severe damage in the brain. The situation becomes fatal when the observed MLS is more than 5 mm (Ono et al., 2001), (Englander et al., 2003). Among the above-listed features, MLS is the prominent feature for SOL and helps determine the patient's survival.

An extensive review (Liao and Chen, 2018) has compared various techniques used to quantify MLS. Automatic MLS detection (Xiao et al., 2010) from CT slices was proposed by decomposing the deformed line segments into three parts where the first and the third parts were the posterior segments representing the straight ideal midline (iML), and the middle part was the part of cerebral hemispheres and was represented by the central curved segment as a quadratic Bezier curve. Parameter optimization (Liao et al., 2010) by a genetic algorithm was implemented to derive the optimal control point values of the Bezier curve. A combination of multi-resolution binary level set method and Hough transform (Xiao et al., 2011) was proposed for the automatic detection of MLS. Detection of anatomical markers such as lateral ventricle using the Gaussian mixture model for the lateral ventricle segmentation, tracing inner boundary contour, and detecting the sharpest turning point for flax detection was proposed (Liu et al., 2014a), which has proved to be useful for detection of MLS. In one study (Chen et al., 2015), enhanced Voigt model, with stress equation modified based on various properties such as the size of the tumor, distance of the tumor from iML, and local symmetry has been applied for the estimation of MLS that was caused by the presence of cerebral gliomas. A method based on the optimization matching template set (Li et al., 2016) was also proposed for the detection of iML. Deep learning framework-based techniques (LeCun et al., 2015), (Greenspan et al., 2016), (Shen et al., 2017) have gained popularity these days for the analysis of medical images. One recent work (Jain et al., 2019) has reported the detection of hematoma using deep learning followed by manual measurement of the midline shift. U-Net architecture is also famous in semantic segmentation of medical images (Zhang et al., 2019), (Rundo et al., 2019), but this approach leads to excessive and redundant use of computational resources as it extracts low-level features.

The steps followed in this current study are summarized in Fig. 2. The left hemisphere was labeled as Class I and the right hemisphere as Class II. The background acted as Class III. The junction of Class I and Class II was the dML. U-net architecture (Ronneberger et al., 2015) was applied to the CT images for hemispheric segmentation, and then dML was traced. Several networks (U-Net, SegNet, and ResNet 50) were trained using our dataset. The accuracy was calculated by the sum of correctly identified pixels upon the total number of pixels multiplied by 100. Computed 5-fold average validation accuracy values of (a) current U-Net structure, (b) SegNet with the architecture encoders with 13 convolutional layers as in (Badrinarayanan et al., 2017), and (c) ResNet 50 with 50 layers, one 3 × 3 convolution (in contrast to two as in (He et al., 2016)) were found to be 94.05 %, 86.74 %, and 85.20 % respectively. For having the highest validation accuracy amongst the tested deep-learning frameworks, the U-Net structure was adopted in this study.

The automated detection of midline shift (MLS) was accomplished using a segmentation approach (segmenting right/left hemispheres). The division of left and right hemispheres was based on the locations of lateral ventricles, the upper flax, and the lower flax. The space between the two lateral ventricles served as the boundary of the two hemispheres, which was learned by our U-Net model. The contextual features such as edges, shapes, contours, etc., computed in the downsampling layers were concatenated to the location information from the upsampling layers with the help of skip connections. This enabled better segmentation by considering both regional and global features. Thus the proposed segmentation approach was more intuitive to the clinicians as it traced the deformed midline at each pixel. A detailed correlation analysis was performed between MLS and clinical severity.

Section snippets

Collection of data

CT data were collected from n = 80 patients visiting one collaborating neuroimaging center (EKO diagnostics, Medical College and Hospitals campus, Kolkata, India) with different types of brain hemorrhages. In our data, it was observed that (a) when the lateral ventricle was replaced by hematoma in intra-ventricular hemorrhage (IVH) cases, clinical finding of midline (iML and dML) was difficult, (b) whereas in the cases with subdural hemorrhages (SDH), no significant midline shift (MLS) was

Methods

The dML curves were extracted from the hemispheric junction. This was followed by skull-stripping and brain-extraction by FSL software (Jenkinson et al., 2012). Then each NCCT slice was segmented into Class I (left hemisphere) and Class II (right hemisphere) using the 2D U-Net model. Ideal and deformed midline (iML and dML) and midline shift (MLS) were then quantified. Finally, MLS indices were correlated with hematoma indices and Radiologist-assigned severity score (RSS) to establish clinical

Midline shift estimation

Few examples of estimated R-L hemispheres, ideal midline, and deformed midlines (predicted and original) are shown in Fig. 5 for comparison. Quantitative errors (from Eqn (1)-(4)) are summarized in Table 1. The tabulated values are the averages across all the folds, except the maximum MLS error, where the value represents the highest MLS error after evaluating across all the folds. Table 2 reports the comparison of actual errors in MLS detection with the existing methods as reported in

Conclusion

This study proposed a novel deep-learning-based approach – to estimate hematoma-affected deformed midline (dML) from neuroimages, even in challenging cases like indistinct brain flax points and brain slices with missing ventricles. The proposed algorithm demonstrated excellent accuracy in dML estimation with low errors in dML tracing (pixel-error = 1.294 mm, area-error = 66.4 mm2, volume-error = 253.73 mm3), significantly better than published state-of-the-art methods. Different predictors

Ethical approval

All procedures performed in studies involving human participants were in accordance with the ethical standard of the institutional and or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Informed consent

Informed consent was obtained from all the participants included in this study.

Declaration of Competing Interest

The authors declare that they do not have any conflict of interest.

Acknowledgment

The first author would like to acknowledge Council for Scientific and Industrial Research (CSIR), New Delhi, India, file number (09/81(1296)/17) for financial support.

References (24)

  • W. Chen et al.

    Actual midline estimation from brain CT scan using multiple regions shape matching

  • M. Chen et al.

    Automatic estimation of midline shift in patients with cerebral glioma based on enhanced voigt model and local symmetry

    Australas. Phys. Eng. Sci. Med.

    (2015)
  • View full text