=Paper=
{{Paper
|id=Vol-2564/shortarticle_5-CRoNe2019
|storemode=property
|title=Convolutional neural networks for detection intracranial hemorrhage in CT images
|pdfUrl=https://ceur-ws.org/Vol-2564/shortarticle_5-CRoNe2019.pdf
|volume=Vol-2564
|authors=Juan Sebastian Castro,Steren Chabert,Carolina Saavedra,Rodrigo Salas
|dblpUrl=https://dblp.org/rec/conf/crone/CastroCSS19
}}
==Convolutional neural networks for detection intracranial hemorrhage in CT images==
https://ceur-ws.org/Vol-2564/shortarticle_5-CRoNe2019.pdf
Proceedings of the 4th Congress on Robotics and Neuroscience
Convolutional neural networks for
detection intracranial hemorrhage
in CT images
Juan Sebastian Castro1 , Steren Chabert1,2 , Carolina Saavedra1 , Rodrigo Salas1,2
*For correspondence:
juan.castro@postgrado.uv.cl (JSC); 1 Universidad de Valparaíso; 2 Centro de Investigación y Desarrollo en Ingeniería en Salud
rodrigo.salas@uv.cl (RS)
Present address: † Escuela de
Ingeniería C. Biomédica,
Universidad de Valparaíso, Chile;
‡ Centro de Investigación y
Abstract Deep learning algorithms have recently been applied for image detection and
Desarrollo en Ingeniería en Salud,
classification, lately with good results in the medicine such as medical image analysis. This paper
CINGS-UV, Universidad de aims to support the detection of intracranial hemorrhage in computed tomography (CT) images
Valparaíso, Chile. using deep learning algorithms and convolutional neural networks (CNN). The motivation of this
work is the difficulty of physicians when they face the task to identify intracranial hemorrhage,
especially when they are in the primary stages of brain bleeding, making a misdiagnosis. A total of
491 CT studies were used to train and evaluate two convolutional neuronal networks in the task of
classifying hemorrhage or non-hemorrhage. The proposed CNN networks reach 97% of recall, 98%
accuracy and 98% of F1 measure.
Introduction
Intracranial hemorrhage (HIC) corresponds to bleeding inside the skull caused by a vascular rupture.
Speed of diagnosis is crucial because the mortality reaches up to 60% after 30 days and 35%
to 52% of patients die before a month after being diagnosed, and approximately half of these
deaths occur within the first 24 hours (Caceres and Goldstein, 2012) (Rodríguez-Yáñez et al., 2013).
This is a reason why HIC is considered a medical emergency and specialists must diagnose it
properly and quickly. However, in general medicine settings and emergency rooms, up to 20%
of patients with suspected HIC may be misdiagnosed, which is an indicator that bleeding cannot
be reliably distinguished without the support of medical imaging techniques (Gross et al., 2019).
Brain neuroimaging computed tomography (CT) for the diagnosis of intracranial hemorrhage, is the
most reliable method during the first week after the onset of HIC. The visualization of intracranial
hemorrhage in CT images depends on density, volume, location, relationship with the surrounding
structures (Cohen, 1992), all previous properties make HIC diagnosis difficult. An automatic process
for HIC detection in the triage workflow, would significantly decrease the time to diagnosis and
expedite treatment.
Automatic or semi-automatic detection of intracerebral hemorrhages in CT images without
contrast is a recent field of research that is follow by advances in artificial intelligence and image
processing. Some of the models proposed for detection are based on K-means and Fuzzy K-means
(Bhadauria et al., 2013) (Zaki et al., 2011) which in some cases are combined with the Otsu method
for segmentation of regions of interest (Loncaric et al., 1999). Other authors propose models based
on the intensity of pixel (Liao et al., 2010), level sets and weights of the histogram (Shahangian
and Pourghassem, 2016), morphological operations (Chan, 2007). On the other hand, in recent
years, the use of deep learning for image classification tasks has become popular, authors present
models using convolutional neural networks (CNN) for the detection of intracranial hemorrhages
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Proceedings of the 4th Congress on Robotics and Neuroscience
(Chilamkurthy et al., 2018) (Helwan et al., 2018), you can also find models that make use of deep
learning for the segmentation of HIC or brain injuries in general (Ito et al., 2019) (Kamnitsas et al.,
2017).
As described, the HIC is classified as a medical emergency in which survival is given by the speed
and effectiveness of the diagnosis. So an algorithm that is used to support the diagnostic task must
be precise and capable of generalizing, in these cases the best results have been obtained using
techniques based on deep learning, that its speed after a previous training of the network. This
work presents the use of convolutional neural networks for the task of classifying hemorrhage vs
non-hemorrhage, 491 studies of computed tomography of the head without contrast were used
with a total of 193,317 slices in which there are 4 types of intracranial hemorrhage and in addition
to brains healthy.
This document is organized as follows. In Section II, the proposed method is presented. The
results are presented in Section III. Discussion of results in Section IV. Finally, the main conclusions
are presented in section V.
Methods and Materials
Data Base
The database chosen is known as CQ500 (Chilamkurthy et al., 2018) and was provided by the
Center for Advanced Research in Imaging, Neurosciences and Genomics (CARING) in New Delhi,
India. This database is part of a set of head CT images taken by several radiologists in the center of
New Dehli. The tomographs used in radiology centers to obtain the images vary between 16 to 128
cuts. The data was taken from the local PACS servers and anonymized according to the internal
guidelines defined in HIPAA. Data were collected in two blocks (B1 & B2). Block B1 was collected
by selecting all CT studies taken at the radiological center for 30 days beginning on November 20,
2017, Block B2 was selected from the remaining studies. Each of the selected studies was evaluated
according to the following exclusion criteria:
1. Patients should not have any post-operative defects such as burr hole/shunt/clips.
2. They should have at least one CT study without axial cut contrast and a soft kernel reconstruc-
tion that contemplates the entire brain
3. Patients should not be less than 7 years old. If age information is not available, it will be
estimated through bone degradation and cranial sutures.
The total of 491 studies were evaluated by three independent expert radiologists with 8, 12 and
20 years experience in the interpretation of cranial CT images. None of the 3 readers participated
in the clinical care or diagnosis of the patients, nor did they have access to their medical history.
Each of the radiologists independently assessed the studies in the CQ500 data set following the
evaluation instructions, the order of presentation of the studies was randomized to minimize
patient remember.
For each CT study the following information was recorded:
• The presence or absence of intracranial hemorrhage, and if its type (intracerebral, subarach-
noid, epidural, subdural), state (chronic or non-chronic) and the affected hemisphere (right,
left) is present
• The presence or absence of midline movement and mass effect
• The presence or absence of fractures. If present, if this is a cranial (partial) fracture
If the three evaluators did not reach a unanimous agreement for each of the studies and
findings, the interpretation of the majority of the evaluators was used as the final diagnosis. The
characteristics of the database used are found in the table 1.
Proceedings of the 4th Congress on Robotics and Neuroscience
Characteristic CQ500 dataset
No. of scans 491 / 193.317 slices
Mean age 22.43
PREVALENCE
No. of scans (percentage) with
205 (41.17%)
intracranial hemorrhage
Intracerebral 162 (32.99%)
Subdural 53 (10.79%)
Extradural 13 (2.64%)
Subarachnoid 60 (12.21%)
Table 1. Dataset characteristics
Image Preprocessing
For the non-contrast CT series original dataset, first, we decided to remove the background image
of all slices, since it does not provide any information to the classification algorithm. Next, instead
of using the entire CT dynamic range, it was decided to windowed the densities of each slice using
a window (level=50, width=80), to visualize only the brain parenchyma. An anisotropic filter was
applied with values (kernel=0.02, time=10) also, the pixel values of each slice in the data set were
normalized between 0 and 1. Finally, resizing to 256 x 256 pixels, all preprocessing was done before
passing to the deep learning models. Figure 1 shows CT slices with the presence of hemorrhage
and no presence after preprocessing stage.
(a) Intracranial Hemorrhage (b) Healthy
Figure 1. Two CT no contrast images after preprocesing
Convolutional Neural Network
Convolutional Neural Network (CNN) is a specialized network in processing grid topology data. The
most common examples are 1-D grid data at regular time intervals, images and 2-D data with pixel
grids. The name for this type of networks arises from the mathematical convolution operation that
the network uses within its processing. In simple words, convolution is the operation between two
functions with a real value argument which is typically denoted as:
𝑆(𝑡) = (𝑥 ∗ 𝑤)(𝑡)
In CNN terminology the first argument (𝑥) is called the input and the second argument (𝑤) is called
the kernel. The output is usually denoted as a feature map (Goodfellow et al., 2016). The principal
function of convolution is feature extraction from input images. Generally, a convolutional neural
network is divided into three stages: first convolutional layers with an activation like ReLU (rectified
linear unit), second a pooling layer for size reduction typically max-pooling. Finally, a flatten layer
before to a fully connected layer to classification the features maps.
Proceedings of the 4th Congress on Robotics and Neuroscience
Figure 2. Proposed Convolutional Neural Network for hemorrhage detection
In this paper, two convolutional neural networks for hemorrhage detection were employed,
first a simple custom CNN was develop maintaining parsimony. The network architecture of the
first CNN is shown in figure 2, start with two convolutional layers with ReLu activation and kernel
(3x3), next a max-pooling layer size reduction, then again two convolutional layers with the same
characteristics than previous convolutional layers followed by a max-pooling layer with kernel (2x2),
next a flatten layer to prepare the features maps to dense layers. Finally, two fully connected
layers for classification was implemented to predict labels with sigmoid activation. We named our
network as CNN4. Another CNN was employed for the hemorrhage detection task, we were decided
to use a popular network VGG16 (Zhang et al., 2015) with a modification for binary classification
(hemorrhage vs no-hemorrhage), VGG16 is one of the most used and reliable CNN tested in a
variety of dataset like ImageNet or CMNIST (Russakovsky et al., 2015), that is compose of 5 blocks
(convolutional + pooling) with 3 fully connected layers used for the classification task.
Training and evaluating models
In this study, the CNN models were trained using the preprocessing slices. A total of 193.317 slices
of 491 CT scans were used to train and evaluate the CNN network. Then, two methods to train the
models were proposed:
1. Slices randomized: All slices were randomized to train (0.85) and test (0.15) sets, regardless
of independence between subjects. This means a part of the slices of a subject could be in
the train set and another part of the slices could be in the test set.
2. Subject randomized: All slices were randomized to train (0.85) and test (0.15) sets, ensuring
independence between subjects. This means all slices of one subject were sent to train or
test.
Due to the need for a large amount of data from the deep learning networks, we decided to divide
the dataset by 0.85 for training, of which 0.2 was used for validation during the training process.
Each model was trained for 150 epochs with a batch size of 32, the best model was saved to be
evaluated with the test set. Binary cross-entropy loss was used to assess performance over time.
Some metrics were obtained to evaluate the performance of CNN in the classification of hemor-
rhage vs non-hemorrhage. Receiver operating characteristic (ROC) curves were obtained for CNN4
and VGG16, with each of the proposed training methods. Accuracy, recall and F1 measure were
also obtained for each of the algorithms.
Results
The ROC curves obtained from the performance evaluation of the VGG16 network (figure 4) show a
much higher performance in the case of the randomized slices (A) method, reaching 0.989 of area
under curve (AUC), as It also presents a recall (table 2) of 0.974 and an F1 measure of 0.971. These
results contrast with those obtained for the method of subject randomized (Figure 4B) where the
AUC is 0.783, with a recall that barely reaches 0.735 and an F1 measure of 0.758, concerning the
Proceedings of the 4th Congress on Robotics and Neuroscience
Figure 3. Architecture of the VGG16 Convolutional Neural Network
accuracy, a 0.707 for the classification of hemorrhage vs no hemorrhage in the test set.
In the case of the CNN4 algorithm, the ROC curves obtained (figure 5) also show a good network
performance for the randomized slices method (figure 5 A), with an area under the curve of 0.982,
as well as a recall of 0.972, F1 measure of 0.972 and an accuracy of 0.981, which are very similar
to those obtained with VGG16. On the other hand, the performance obtained with the subject
randomized method in the CCN4 network were: AUC 0.658, recall 0.721, F1 measure 0.687 and
accuracy of 0.598, as well as their respective ROC curve (figure 5 B). All the performance metrics
obtained for the test set can be found in table 2.
(a) VGG16 trained with slices randomized (b) VGG16 trained with subject randomized
Figure 4. ROC curves for VGG16 network trained with the two proposal methods for detection intracranial
hemorrhage. Area Under Curve (AUC) is also presented
(a) CNN4 trained with slices randomized (b) CNN4 trained with subject randomized
Figure 5. ROC curves for CNN4 network trained with the two proposal methods for detection intracranial
hemorrhage. Area Under Curve (AUC) is also presented
Discussion
The ROC curves obtained in figure 4 and figure 5 for both the VGG16 network and the proposed
CNN4 network, present excellent results for the classification of hemorrhage vs. non-hemorrhage
using the training method of slices randomized, reaching 0.98 of AUC in both cases. This represents
Proceedings of the 4th Congress on Robotics and Neuroscience
Area Under
Mode Accuracy Recall F1_measure
Curve (ROC)
VGG-16
Slices Randomized 0.968 0.974 0.971 0.989
Subject Randomized 0.707 0.735 0.758 0.783
CNN - 4
Slices Randomized 0.981 0.972 0.982 0.982
Subject Randomized 0.598 0.721 0.687 0.658
Table 2. Performance of algorithms for CQ500 dataset
a high recall and specificity of both algorithms, this is confirmed with the f1 measure that in both
cases exceeds 0.97, this metric is the compromise between recall and accuracy. This performance is
explained by the nature of the images used for the classification of hemorrhage vs non-hemorrhage.
A CT scan is composed of several slices that represent the skull in this case in different positions
in a cut (axial for this study), as all slices belong to the same patient contain similarities. On the
other hand, there is no difference between the performance of the two classification algorithms
proposed with the randomized slices method, the results show that the CNN4 network despite
being simpler than the VGG16 can have a similar performance.
The results presented for the subject randomized method in figure 4 and figure 5 for both
VGG16 and CNN4 networks do not present the expected performance. Table 2 shows the metrics
obtained, which highlights the recall that exceeded 0.72 for both networks evaluated, the per-
formance obtained by the VGG16 being outstanding, which in this case exceeds more than 10
percentage points in the metrics of AUC and F1 measure to the CNN4 network. So for this subject
randomized training method, the VGG16 network has better results than CNN4. On the other hand,
although the performance of both VGG16 and CNN4 networks was not as expected, the subject
randomized method has the advantage of preserving the independence of the data and therefore
having a better capacity to generalize.
Regarding the performance of the networks compared to the state of the art, it can be deter-
mined that the randomized slices method with a recall of 0.974 for the VGG16 and 0.972 for CNN4,
is very similar to the performance obtained by other studies (Chilamkurthy et al., 2018) (Helwan
et al., 2018). And although the performance obtained with the subject randomized method does
not reach these recall levels, it still exhibits outstanding performance over traditional medical image
processing techniques (Bhadauria et al., 2013) (Zaki et al., 2011).
Conclusion
In this paper, two convolutional neural networks were proposed for the task of classification
of intracranial hemorrhage vs. non-hemorrhage, a popular VGG16 network, and a CNN4 own
network. Two different training methods were also proposed (slices randomized and subject
randomized) where the second ensures the independence of the data. The results show an
outstanding performance for the first training method in the classification task, on the other
hand, the second training method proposed is at the level of the classic medical image processing
techniques. With this, it can be concluded that convolutional neural networks are a useful tool for
the identification of intracranial hemorrhages in computed tomography images and can be used as
a support in the diagnosis of this type of pathologies. Additionally, it was found that the method for
choosing the train set and test set is influential for the performance of deep learning algorithms.
Therefore, a greater study of the independence of the data in the use of computed tomography
images is required for classification through convolutional neural networks.
Proceedings of the 4th Congress on Robotics and Neuroscience
References
Bhadauria HS, Singh A, Dewal ML. An integrated method for hemorrhage segmentation from brain CT Imaging.
Computers & Electrical Engineering. 2013 Jul; 39(5):1527–1536. https://linkinghub.elsevier.com/retrieve/pii/
S0045790613000955, doi: 10.1016/j.compeleceng.2013.04.010.
Caceres JA, Goldstein JN. Intracranial Hemorrhage. Emergency Medicine Clinics of North Amer-
ica. 2012 Aug; 30(3):771–794. https://linkinghub.elsevier.com/retrieve/pii/S0733862712000272, doi:
10.1016/j.emc.2012.06.003.
Chan T. Computer aided detection of small acute intracranial hemorrhage on computer tomography of brain.
Computerized Medical Imaging and Graphics. 2007 Jun; 31(4-5):285–298. https://linkinghub.elsevier.com/
retrieve/pii/S0895611107000183, doi: 10.1016/j.compmedimag.2007.02.010.
Chilamkurthy S, Ghosh R, Tanamala S, Biviji M, Campeau NG, Venugopal VK, Mahajan V, Rao P, Warier P.
Development and Validation of Deep Learning Algorithms for Detection of Critical Findings in Head CT Scans.
arXiv:180305854 [cs]. 2018 Mar; http://arxiv.org/abs/1803.05854, arXiv: 1803.05854.
Cohen W. Computed tomography of intracranial hemorrhage. Radiologic Clin North Amer. 1992; 2:75–87.
Goodfellow I, Bengio Y, Courville A. Deep Learning. MIT Press; 2016. http://www.deeplearningbook.org.
Gross BA, Jankowitz BT, Friedlander RM. Cerebral Intraparenchymal Hemorrhage: A Review. JAMA.
2019 Apr; 321(13):1295. http://jama.jamanetwork.com/article.aspx?doi=10.1001/jama.2019.2413, doi:
10.1001/jama.2019.2413.
Helwan A, El-Fakhri G, Sasani H, Uzun Ozsahin D. Deep networks in identifying CT brain hemorrhage. Journal of
Intelligent & Fuzzy Systems. 2018 Aug; 35(2):2215–2228. http://www.medra.org/servlet/aliasResolver?alias=
iospress&doi=10.3233/JIFS-172261, doi: 10.3233/JIFS-172261.
Ito R, Nakae K, Hata J, Okano H, Ishii S. Semi-supervised deep learning of brain tissue segmentation. Neu-
ral Networks. 2019 Aug; 116:25–34. https://linkinghub.elsevier.com/retrieve/pii/S0893608019300954, doi:
10.1016/j.neunet.2019.03.014.
Kamnitsas K, Ledig C, Newcombe VFJ, Simpson JP, Kane AD, Menon DK, Rueckert D, Glocker B. Effi-
cient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Medical Im-
age Analysis. 2017 Feb; 36:61–78. https://linkinghub.elsevier.com/retrieve/pii/S1361841516301839, doi:
10.1016/j.media.2016.10.004.
Liao CC, Xiao F, Wong JM, Chiang IJ. Computer-aided diagnosis of intracranial hematoma with brain deformation
on computed tomography. Computerized Medical Imaging and Graphics. 2010 Oct; 34(7):563–571. https:
//linkinghub.elsevier.com/retrieve/pii/S0895611110000388, doi: 10.1016/j.compmedimag.2010.03.003.
Loncaric S, Dhawan AP, Cosic D, Kovacevic D, Broderick J, Brott T. Quantitative intracerebral brain hemorrhage
analysis. In: Hanson KM, editor. San Diego, CA; 1999. p. 886–894. http://proceedings.spiedigitallibrary.org/
proceeding.aspx?articleid=980754, doi: 10.1117/12.348648.
Rodríguez-Yáñez M, Castellanos M, Freijo MM, López Fernández JC, Martí-Fàbregas J, Nombela F, Simal P, Castillo
J, Díez-Tejedor E, Fuentes B, Alonso de Leciñana M, Álvarez Sabin J, Arenillas J, Calleja S, Casado I, Dávalos A,
Díaz-Otero F, Egido JA, Gállego J, García Pastor A, et al. Guías de actuación clínica en la hemorragia intracerebral.
Neurología. 2013 May; 28(4):236–249. https://linkinghub.elsevier.com/retrieve/pii/S0213485311001447, doi:
10.1016/j.nrl.2011.03.010.
Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A, Bernstein M, et al.
Imagenet large scale visual recognition challenge. International journal of computer vision. 2015; 115(3):211–
252.
Shahangian B, Pourghassem H. Automatic brain hemorrhage segmentation and classification algorithm based
on weighted grayscale histogram feature in a hierarchical classification structure. Biocybernetics and Biomed-
ical Engineering. 2016; 36(1):217–232. https://linkinghub.elsevier.com/retrieve/pii/S0208521615300036, doi:
10.1016/j.bbe.2015.12.001.
Zaki WMDW, Fauzi MFA, Besar R, Ahmad WSHMW. Abnormalities detection in serial computed tomography
brain images using multi-level segmentation approach. Multimedia Tools and Applications. 2011 Aug;
54(2):321–340. http://link.springer.com/10.1007/s11042-010-0524-0, doi: 10.1007/s11042-010-0524-0.
Zhang X, Zou J, He K, Sun J. Accelerating very deep convolutional networks for classification and detection. IEEE
transactions on pattern analysis and machine intelligence. 2015; 38(10):1943–1955.