You are on page 1of 6

Comparative Study on Stroke Lesion Core

Segmentation in CTP Images


1st Amira Esawy 2rd Sahar Fawzi 3rd Mustafa Elattar
Medical Imaging and Medical Imaging and Medical Imaging and
Image Processing Group Image Processing Group Image Processing Group
Center of Informatics Science Center of Informatics Science Center of Informatics Science
Nile University Nile University Nile University
Giza, Egypt Giza, Egypt Giza, Egypt
aesawy@nu.edu.eg sfawzi@nu.edu.eg melattar@nu.edu.eg

Abstract—Automatic segmentation of ischemic stroke lesion However, it cannot be widely used in the clinical setting
using Computed Tomography Perfusion (CTP) images becomes due to limitations in availability, time, and cost. Compared
an essential tool for quantification of the extension of stroke in to MRI, CT scans are rapid, cost-effective, widely available,
early stages because of its availability, speed, and low cost. In this
paper, we combined non-contrast CT scans which help in showing and they are already used by clinicians as an important first
the core of the lesion along with CTP perfusion maps that show step in the triage, diagnosis, and treatment assessment of acute
the abnormality of perfusion to enhance the core segmentation ischemic stroke [2]. CTP (computed tomography perfusion) is
task. We utilized the performance of Unet, ResUnet, and Re- used to assess blood perfusion in the brain. To obtain CTP
sUnet++ architectures on the ISLES 2018 dataset in order to use images, an intravenous contrast agent is first injected, followed
a single architecture for direct segmentation. Data augmentation
helped in increasing the number of training samples, overcome by repeated scans as it spreads through the brain. While
overfitting issues, and teaching the network the invariance and CT shows the core of the lesion, CTP shows all areas with
robustness properties. Unet with data augmentation achieved a abnormal perfusion, including the core and penumbra. The
dice coefficient equal to 0.65 while ResUnet++ achieved a dice combination of both is also speedy to obtain and may provide
coefficient equal to 0.55 without using data augmentation. enough reliable data for automatic analysis. Furthermore, CT
Index Terms—Medical image analysis, CT, CT perfusion,
Automatic lesion segmentation, Convolutional neural networks
perfusion has been used to triage patients with acute stroke
instead of MRI, which may reduce scanning time for stroke
patients [3]. A series of recent studies have indicated that
deep learning technology, – especially Convolutional Neural
I. I NTRODUCTION
Network (CNN), has achieved significant success in a variety
According to World Health Organization (WHO), Stroke of computer vision tasks such as classification [4], detection
is the second global cause of death and the third cause of [5], and segmentation. Also, CNN has proved its power and
disability-adjusted life years. 15 million people suffer from effectiveness in biomedical segmentation tasks. Criesan et
stroke yearly, 5 million of them die and another 5 million al. [6] applied CNN to a medical image segmentation task,
suffer from disabilities.. Stroke is a fatal disease that happens predicting a pixel’s label based on information in a separate
when the blood supply to a portion of the brain is stopped. square window around it. Later, the fully convolutional net-
This leads to a part of the brain cannot get the nutrition work (FCN) [7] is proposed, which can predict the image’s
it needs and cause brain cells to die. It has two types [1], pixel-label in a single step forward operation. UNet [8] is
ischemic stroke happens when a blood vessel supplying the an FCN-based method for combining localization and context
brain is obstructed. It is about 87 percent of all strokes. While information using an encoder-decoder structure and skip-
hemorrhagic stroke occurs due to the rupture of a weakened connections. Nielsen et al. [9] pioneered the CNN method
blood vessel. The stroke lesion is divided into two areas, the for stroke segmentation using a simple deep encoder-decoder
infarct core, irreversibly damaged tissue, and the penumbra, structure. Recently, Deep Learning began to be applied to CT
tissue at risk that can be recovered if blood flow is restored scans for stroke lesion segmentation in the 2018 edition of the
quickly. The localization and quantification of the acute core or Ischemic Stroke Lesion Segmentation (ISLES) challenge. This
penumbra are considered a crucial clinical interest because it competition began in 2015 to provide a venue for a fair and
helps in evaluating the amount of tissue that can be recovered direct comparison of automated stroke imaging approaches. In
with different treatments and make more informed decisions. 2018, the fourth edition released the first public acute stroke
Brain imaging modalities based on Magnetic resonance images dataset based on CT and CTP images. Four of the five compe-
(MRI) and Computed Tomography (CT) are the most effective tition finalists employ CNNs based on the U-Net architecture.
modalities used to assess and evaluate stroke lesions. MRI Song et al. [10], and P. Liu et al. [11] proposed a CTP data
is more sensitive to infarction’s early parenchymal changes. analysis framework using Generative Adversarial Networks
(GAN) [12] to generate Diffusion-Weighted Imaging (DWI)
and segment the region of stroke they achieved an average
dice coefficient of 62.4%, 60.65% respectively.
While Clerigues et al. [13] introduced in their research an
automated tool based on an asymmetrical residual encoder-
decoder CNN architecture for acute stroke segmentation using
ISLES 2018 dataset and achieved a Dice similarity coefficient
of 49%. The approach by Chen et al. [14] managed to achieve
a dice coefficient equal to 48% by using the ensembling
approach of multiple models.
In this work, We proposed a segmentation technique by a
single network.
• Using non-contrast CT raw images concatenated with
CTP perfusion maps.
• Compare the performance of three CNN architectures
used popularly for semantic segmentation tasks on ISLES
2018 dataset, U-net, ResUnet [15], ResUnet++ [16].
• Study the effect of using data augmentation on the
performance of the three networks and increasing the
number of training samples.

II. M ETHODS
A. Dataset and Data Preprocessing
The Framework is trained and tested using ISLES 2018
challenge. This year’s challenge offers acute stroke CT perfu-
sion imaging scans as well as manually delineated core lesions
on MRI DWI images obtained shortly afterward. The dataset
provided contains 94 labeled training images and 62 unlabeled
testing images. The images were captured as slabs with a
variable number of axial slices ranging from 2 to 22 depending
on the patient, with a 5 mm spacing and a resolution of 256
*256. In training, each case has eight modalities or parametric
maps, including CT, CT 4DPWI, CT MTT, CT Tmax, CT
CBF, CT CBV, MR DWI, and ground truth, OT. CT MTT, CT
Tmax, CT CBV, and CT CBF channels are concatenated with
the non-contrast CT scan during the training phase. In addition,
when feeding training samples, we perform data augmentation.
Augmentation operations such as scaling, flipping, rotation, Fig. 1. Unet Architecture
and adding Gaussian filters are used.

B. Architectures
a) Unet: Unet architecture proposed by [8] is a powerful b) ResUnet: ResUnet [15] is a modified architecture
CNN model for medical imaging applications. It is a U-shaped from U-Net that has shown state-of-the-art results for road
structure made up of a contracting path (on the left side) and image extraction. It is a semantic segmentation neural network
an expansive path (on the right side). The contracting path, that is built up by combining the strengths of both U-Net and
considered an encoder, is a convolutional network composed of residual neural networks. Fig. 2 illustrates the architecture of
repeated convolutional layers, each followed by a rectified lin- the network. The benefits of this combination are that the
ear unit (ReLU) and a max-pooling operation. In the encoder, residual unit will make the network’s training easy, and the
spatial information is reduced while feature information is presence of the skip connections within a residual unit and
increased. The decoder in the expansive path combines feature between low and high levels of the network will overcome the
and spatial information through a series of up-convolutions problem of degradation and facilitate information propagation.
and concatenations with high-resolution features from the This would design a neural network with much fewer param-
contracting path. Fig. 1 illustrates the architecture of Unet. eters however could achieve better performance on semantic
segmentation.
in natural language processing (NLP). It has the ability to
determine which parts of the network need attention, this
would help in reducing the computational of the encoding and
boosts the results.

Fig. 2. ResUnet Architecture

c) ResUnet++: The ResUnet ++ architecture [16] is a


modified network of ResUnet taking the benefits of using
residual blocks, the squeeze and excitation block, Atrous
Spatial Pyramidal Pooling (ASPP), and the attention block.
It is consists of one stem block then three encoder blocks,
ASPP, and three decoder blocks as Fig. 3 illustrates. The Fig. 3. ResUnet++ Architecture
residual blocks compose of batch normalization, Rectified
Linear Unit (ReLU) activation, and convolutional layers which
III. E XPERIMENT
facilitate the training process by decreasing the parameters
and enhance the model performance, while the squeeze and A. Implementation details
excitation block can help in increasing the network sensitivity All architectures were implemented using Keras with Ten-
in relevant features and remove unnecessary ones. ASSP is a sorflow and ran on Google Colaboratory. We concatenate raw
bridge between the encoder and decoder paths that controls the CT scans with CT perfusion parameters and fed them to
field of view for capturing multi-scale information at different the model. We split the training set into 80% and 20% for
rates which have shown promising results on the semantic validation, we used this portion to evaluate the model since the
segmentation tasks. The attention block has been widely used test set provided was unlabeled. All images have the same size
256*256 and Adam optimizer was used with a 0.001 learning TABLE III
rate. The model was trained for 100 epochs. C OMPARISON WITH THE TOP FOUR METHODS FOR ISLES 2018 TESTING
SET
B. Results
Method DC
a) Experiment A: Firstly, we trained the three archi- Ours 0.65
tectures ResUnet++, ResUnet, and Unet with the training Song et al. [10] 0.62
Liu et al. [11] 0.60
samples and evaluated them using the validation samples. Clerigues et al. [13] 0.49
After concatenation, we had 428 samples for training and Chen et al. [14] 0.48
74 for validation. The input size is (B, 256, 256, 5) where
B is the batch size. The model loss was assessed using
dice loss and dice coefficient (DC), MeanIOU, precision, and
Recall used as evaluation matrix. Hyper-parameters tuning was is the key to our experiments. It increases the informative
the most important step in our experiment. We have tried data about the stroke and improves the results comparing
manually different parameters’ sets (i.e., learning rate, number to direct segmentation using only CTP maps. We compare
of epochs, and batch size) through many experiments and the performance of three networks using the concatenated
evaluating their results until we reached the best results. data with the same settings. In the first experiment, with low
Table I shows the results of ResUnet++, ResUnet and Unet input data ResUnet++ shows the best average dice coefficient
after data augmentation. Fig. 5 shows the predicted lesion core over the testing samples equal to 55% thanks to using the
using the best achieved model. residual units, the squeeze and excitation block, ASPP, and
the attention block. The small number of training samples
TABLE I inspires us to perform the second experiment using data aug-
E VALUATION METRICS ON ISLES 2018 WITHOUT DATA AUGMENTATION mentation techniques. Data augmentation helps in increasing
the training samples and teaching architectures the invariance
Method DC MeanIOU Precision Recall
ResUnet++ 0.55 0.6 0.71 0.45 and robustness properties. The results are increased in the three
ResUnet 0.45 0.69 0.78 0.46 networks compared to the first experiment and Unet shows the
Unet 0.43 0.72 0.80 0.45 best average dice coefficient over the testing samples equal
to 65%. We have a novel segmentation technique using a
b) Experiment B: As we have small number of training single network with simple training procedures and achieve
data, we performed some data augmentation techniques to high results on ISLES 2018 dataset. We succeed to achieve
increase the training samples fed to the model and compare the a dice coefficient higher than the top five methods of ISLES
effect of the data augmentation between the three architectures. 2018 leaderboard. The approach used by Song et al., the first
Data augmentation applied where horizontal and vertical flip, ranked on the challenge, achieved a dice coefficient equal to
rotation with angles 45, 90, and 180, scaling with 0.5 factor 62.4% by GANs to generate DWI images from CTP maps
and adding Gaussian filter. We performed them on the training for segmentation which increases the memory, processing time
samples. They increased more than before by nearly 7%, and need complex training procedures. Moreover, GANs suffer
we had 2996 training samples. We used the same hyper- from non-convergence of parameters and diminished gradient
parameters in experiment A. problems.
Table II shows the results of ResUnet++, ResUnet and Unet However achieving a high DC score, We will need to study
after data augmentation Fig. 4 shows the predicted lesion core the effect of changing the batch size and use the method of
using the best achieved model. ensembling different models.

TABLE II V. C ONCLUSION
E VALUATION METRICS ON ISLES 2018 WITH DATA AUGMENTATION

Method DC MeanIOU Precision Recall In this work, we proposed a direct framework for ischemic
ResUnet++ 0.63 0.71 0.75 0.60 stroke lesion segmentation using non-contrast CT scans and
ResUnet 0.62 0.73 0.76 0.62 CTP perfusion maps and make a comparative study between
Unet 0.65 0.75 0.73 0.70
Unet, ResUnet, and ResUnet++ on ISLES 2018 dataset. We
enhance the performance of the three networks using data
Compared to ISLES 2018 leaderboard [17], our methods augmentation. This framework can be a quick diagnostic
achieved a high score of dice coefficient using U-net with tool for ischemic stroke using CTP images with the limited
data augmentation as Table III shows. availability of MRI images.
Fig .4 shows the predicted lesion core using the best achieved
model.
R EFERENCES
IV. D ISCUSSION
[1] C. Sudlow and C. Warlow, “Comparable studies of the incidence of
Due to the low resolution of CTP perfusion parameters, stroke and its pathological types: results from an international collabo-
concatenation between the CT scans along with CTP maps ration,” Stroke, vol. 28, no. 3, pp. 491–499, 1997.
Fig. 4. Lesion core segmentation using U-net with data augmentation

Fig. 5. Lesion core segmentation using ResUnet++ without data augmentation

[2] M. H. Lev, J. Farkas, J. J. Gemmete, S. T. Hossain, G. J. Hunter, W. J. information processing systems, vol. 28, pp. 91–99, 2015.
Koroshetz, and R. G. Gonzalez, “Acute stroke: improved nonenhanced ct [6] D. Ciresan, A. Giusti, L. Gambardella, and J. Schmidhuber, “Deep neural
detection—benefits of soft-copy interpretation by using variable window networks segment neuronal membranes in electron microscopy images,”
width and center level settings,” Radiology, vol. 213, no. 1, pp. 150–155, Advances in neural information processing systems, vol. 25, pp. 2843–
1999. 2851, 2012.
[3] Y. Yu, Q. Han, X. Ding, Q. Chen, K. Ye, S. Zhang, S. Yan, B. C. [7] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks
Campbell, M. W. Parsons, S. Wang et al., “Defining core and penumbra for semantic segmentation,” in Proceedings of the IEEE conference on
in ischemic stroke: a voxel-and volume-based analysis of whole brain computer vision and pattern recognition, 2015, pp. 3431–3440.
ct perfusion,” Scientific reports, vol. 6, no. 1, pp. 1–7, 2016. [8] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks
[4] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification for biomedical image segmentation,” in International Conference on
with deep convolutional neural networks,” Advances in neural informa- Medical image computing and computer-assisted intervention. Springer,
tion processing systems, vol. 25, pp. 1097–1105, 2012. 2015, pp. 234–241.
[5] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time [9] A. Nielsen, M. B. Hansen, A. Tietze, and K. Mouridsen, “Prediction
object detection with region proposal networks,” Advances in neural of tissue outcome and assessment of treatment effect in acute ischemic
stroke using deep learning,” Stroke, vol. 49, no. 6, pp. 1394–1401, 2018.
[10] T. Song, “Generative model-based ischemic stroke lesion segmentation,”
arXiv preprint arXiv:1906.02392, 2019.
[11] P. Liu, “Stroke lesion segmentation with 2d novel cnn pipeline and
novel loss function,” in International MICCAI Brainlesion Workshop.
Springer, 2018, pp. 253–262.
[12] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,”
Advances in neural information processing systems, vol. 27, 2014.
[13] A. Clerigues, S. Valverde, J. Bernal, J. Freixenet, A. Oliver, and
X. Lladó, “Acute ischemic stroke lesion core segmentation in ct per-
fusion images using fully convolutional neural networks,” Computers in
biology and medicine, vol. 115, p. 103487, 2019.
[14] X. Hu, W. Huang, S. Guo, and M. R. Scott, “Strokenet: 3d local
refinement network for ischemic stroke lesion segmentation,” in Int.
MICCAI Brainlesion Workshop, 2018.
[15] Z. Zhang, Q. Liu, and Y. Wang, “Road extraction by deep residual u-
net,” IEEE Geoscience and Remote Sensing Letters, vol. 15, no. 5, pp.
749–753, 2018.
[16] D. Jha, P. H. Smedsrud, M. A. Riegler, D. Johansen, T. De Lange,
P. Halvorsen, and H. D. Johansen, “Resunet++: An advanced archi-
tecture for medical image segmentation,” in 2019 IEEE International
Symposium on Multimedia (ISM). IEEE, 2019, pp. 225–2255.
[17] “Challenges ” isles challenge 2018,”
https://www.smir.ch/ISLES/Start2018.

You might also like