DEEP LEARNING BASED ISCHEMIC STROKE

CLASSIFICATION

U .V. RAM MOHAN RAO(19341A05H1)

R .SAI MANIKANTA(19341A05E2)
V.DURGA PRASAD(19341A05H9)
S .MODHA MADHAN(20345A0516)
Y. VENKATA SAI KALYAN(19341A05J0)

Under the Guidance of

Dr. K. Lakshmana Rao
Associate Professor
Dept. of Computer Science & Engineering
GMR INSTITUTE OF TECHNOLOGY, Rajam
 ABSTRACT
Identifying the blood clot in the brain and predicting the type of stroke plays a vital role in the diagnosis and treatment of
various cerebrovascular diseases. Ischemic stroke is the most common type of stroke and it is developed due to the blockage in
the arteries around the brain. Paralysis, memory loss, change in behavior etc. are most common symptoms which occurs due to
ischemic stroke. Several types of strokes are there in the existence, few of them are Large Artery Atherosclerosis (LAA) and
Cardioembolic (CE). Identifying the type of the stroke will play a vital role in diagnosing the patients. Artificial Intelligence
models like Machine learning and Deep learning based convolution neural networks techniques can be used to address this
problem. Dataset consists of high-resolution blood clot images in .tiff (Tagged Image File Format). Features are extracted from
the images using several preprocessing techniques. With the help of transfer learning, the pre-trained EfficientNet is used to
predict the type of the stroke (CE or LAA). Model performance will be evaluated using the accuracy, precision and recall
metrics.

Keywords: Ischemic Stroke, Deep Learning, Machine learning, Convolutional Neural ,  Image Preprocessing, EfficientNet.
 INTRODUCTION
Stroke is the second leading cause of death worldwide, the typical symptom of acute stroke is the sudden onset of a
focal neurologic deficit, such as dysphasia, hemianopia, sensory loss, etc. These symptoms may develop into
chronic diseases (e.g. dementia, hemiplegia, etc.), which can profoundly affect patients life and consume a large
part of social healthcare cost.

The Ischemic stroke is due to the loss of blood supply to an area of the brain. It is a common type of
stroke. Hemorrhagic stroke is due to bleeding into the brain by the rupture of a blood vessel.
Stroke is divided into two types, one is cerebral vascular obstruction caused by ischemic stroke, and the other is
the cerebral blood vessels burst bleeding caused by haemorrhagic stroke. No matter which type of stroke will
cause abnormal brain function, leading to loss of local function or even brain tissue necrosis.

In detection of ischemic stroke many researches had done on brain tissues some of the techniques are MRI,CT
scan etc. An MRI uses powerful radio waves and a magnetic field to create a detailed view of the brain.
An MRI can detect brain tissue damaged by an ischemic stroke and brain haemorrhagic. In general, before carrying
out special treatment of stroke, patients will first have CT or MRI, after the treatment the doctor will then decide
according to the results. CT images are classified and then used Data Augmentation method to increase the number
of patches, and finally input to convolutional neural network (CNN) for training and testing.
In our project we are using Mayo Clinic Strip-Ai dataset to detect the ischemic stroke region. The dataset consists of
735 images with higher resolution which are in .tiff format. In every image of the dataset consists of
Unwanted whitespaces, by these gaps the image resolution and size is high. We removed the whitespaces in the images
then we apply pre-processing techniques to the images. 
After removing whitespaces from the images we have to resize the image into a fixed format by considering various
labels, such as CE and LAA are the two labels are used. Resizing is useful to reduce the error rate and applying the Hog
model to extract features on specified set of rules.
Data Augmentation is used to balance the imbalanced dataset. In our dataset we have 500 CA Labels and 200 LAA
labels which are not sufficient to our project so we apply data over sampling technique to increase the images up to
1000. The over sampling is achieved by
Various ways such as Random Over Sampling, Smote, K Means Smote. We are using smote technique to balance the
dataset.
After Pre-processing the data we are using hog model to extract the features. HOG, or Histogram of Oriented Gradients,
is a feature descriptor that is often used to extract features from image data. It is widely used in computer vision tasks
for object detection.
The HOG descriptor focuses on the structure or the shape of an object. HOG is able to provide the edge direction as
well. This is done by extracting the gradient and orientation (or you can say magnitude and direction) of the edges.
After feature extraction we the data to ML or DL model to detect the clot region in the images and apply testing to the
test data then verify accuracy to our model.
 LITERATURE SURVEY
[1].Zhang, R., Zhao, L., Lou, W., Abrigo, J. M., Mok, V. C., Chu, W. C., ... & Shi, L. (2018). Automatic segmentation of acute
ischemic stroke from DWI using 3-D fully convolutional DenseNets. IEEE transactions on medical imaging, 37(9), 2149-2160.

•  Ischemic stroke is the most common type of stroke and it is developed due to the blockage in the arteries around the brain.

• A novel automatic method is used to segment acute ischemic stroke from diffusion weighted images (DWIs) using deep 3-D
convolutional neural networks (CNNs).

• Integrating dense connectivity to boost information and gradient propagation within the deep 3D model, and optimize it with
Dice objective function to tackle the severe class imbalance problem.

• A DWI data set containing 242 subjects (90 for training, 62 for validation, and 90 for testing) with various types of acute
ischemic stroke was constructed to evaluate the method.

• This paper achieved high performance on various metrics (Dice similarity coefficient: 79.13%, lesion wise precision: 92.67%,
and lesion wise F1 score: 89.25%.
[2].C. -L. Chin et al., "An automated early ischemic stroke detection system using CNN deep learning algorithm," 2017 IEEE
8th International Conference on Awareness Science and Technology (iCAST), 2017, pp. 368-372, doi:
10.1109/ICAwST.2017.8256481.

• A convolutional neural network (CNN) is a class of deep , feed-forward artificial neural network that have successfully   applied to
analyzing visual imagery. CNNs use a variation of multilayer perceptrons designed to require minimal preprocessing.

• Select the patch images which are used  Data Augmentation method to increase the number of patch images, then input the
patch images into the convolutional neural network for training and testing.

• The input patch image size is 32×32, and the filter size in each convolutional layer is 5×5, and each convolutional layer includes a
ReLU (Rectified Linear Unit) nonlinearity as activation function.

• The CNN module which consists two convolutional layers, one max pooling layer and a single fully connected layer.

• In CNN, the backpropagation learning algorithm is adapted to calculate the error between the actual output and the expected output,
and the network finds the local optimal solution to adjust the weights of each layer by Gradient Descent.

• The performance index includes the Accuracy, TPR, FNR, FPR, and TNR.
[3].K. C. Ho, W. Speier, H. Zhang, F. Scalzo, S. El-Saden and C. W. Arnold, "A Machine Learning Approach for Classifying
Ischemic Stroke Onset Time From Imaging," in IEEE Transactions on Medical Imaging, vol. 38, no. 7, pp. 1666-1676, July
2019, doi: 10.1109/TMI.2019.2901445.

• Demonstrating a machine learning approach for (Time Series Stroke) TSS classification using routinely acquired imaging
sequences.

• To develop imaging features from the magnetic resonance (MR) images and train machine learning models to classify the TSS.

• A deep-learning model is also used to extract hidden representations for the MR perfusion-weighted images and demonstrate
classification improvement by incorporating these additional deep features.

• A correlation method is used to interpret the deep AE features and demonstrated that our proposed classification method is
robust to variations in imaging acquisition.

• The cross-validation results show that our best classifier achieved an area under the curve of 0.765, with a sensitivity of 0.788
and a negative predictive value of 0.609, outperforming existing methods.
[4] Kamal H, Lopez V, Sheth SA. Machine Learning in Acute Ischemic Stroke Neuroimaging. Front Neurol. 2018 Nov
8;9:945. doi: 10.3389/fneur.2018.00945. PMID: 30467491; PMCID: PMC6236025.

• The evaluation and treatment of Acute Ischemic Stroke (AIS) have experienced a significant advancement over the past few
years, increasingly requiring the use of neuroimaging for decision-making.

• Researchers developed a deep learning algorithm based on an autoencoder architecture to extract imaging features in
perfusion-weighted images (PWI) in MRI to determine the time elapsed since stroke onset.

•  A predictive model was introduced for seizure occurrence in a single patient. This approach was based on a multi-class
support vector machine (SVM) and 14 selected features of an electroencephalogram in patients with epilepsy

• A novel training scheme was proposed based on two main components, a 3D CNN that produces accurate soft segmentation
maps and a connected Conditional Random Field that imposes regularization constraints on the CNN output and produces
the final hard segmentation labels.

• Deep learning algorithms using medical imaging often require datasets of tremendous magnitude, the types of which may
not be readily available.

• The predicted time of seizure with a window between 20 and 25 min was reported with an average sensitivity of 90.15,
99.44% specificity, and 97% accuracy.
[5] Kim C, Zhu V, Obeid J, Lenert L. Natural language processing and machine learning algorithm to identify brain
MRI reports with acute ischemic stroke. PLoS One. 2019 Feb 28;14(2):e0212778. doi: 10.1371/journal.pone.0212778.
PMID: 30818342; PMCID: PMC6394972..

• Stroke is one of the leading causes of death and morbidity worldwide and a major health problem according to the Global
Burden of Disease study.

• Institutional Review Boards and Ethics Committee at Chuncheon Sacred Heart Hospital had stored the MRI reports of the
patients nearly 11000 scanned reports as the dataset.

• To classify the two reference standards of acute ischemic stroke (AIS) and non-AIS, four types  ML algorithms—binary
logistic regression (BLR), naïve Bayesian classification (NBC), single decision tree (SDT) and support vector machine
(SVM).

• In this paper by using support vector machine (SVM) using n-gram technique obtained lesser F1 score so to improve the
measure applied bag of word technique to add additional features.

• The single decision tree had  highest accuracy of 98.0% and the highest F1 measure of 93.2 .
[6] Karthik, R., Gupta, U., Jha, A., Rajalakshmi, R., & Menaka, R. (2019). A deep supervised approach for ischemic
lesion segmentation from multimodal MRI using Fully Convolutional Network. Applied Soft Computing, 84, 105685.

• This paper mainly deals with the ischemic lesion segmentation from multimodal MRI images. The authors presented a
fully convolutional network (FCN) to segment the ischemic stroke lesions from the MRI images.

• Authors used the benchmark MRI datasets from the ISLES 2015 challenge. The datasets were provided in the
Neuroimaging informatics technology initiative (NIfTI) format.

• Data preprocessing and Data Augmentation techniques are used to prepare the dataset for further tasks.

• The existing U-Net architecture was enhanced by the author with the leaky Relu functions at the end of two layers to
preserve significant high-level features. 

• Hyperparameter tuning is performed for the loss function, activation and optimization functions which achieved a dice
coefficient of 0.70.
[7] Van Os, H. J., Ramos, L. A., Hilbert, A., Van Leeuwen, M., Van Walderveen, M. A., Kruyt, N. D., ... & Mr Clean
Registry Investigators. (2018). Predicting outcome of endovascular treatment for acute ischemic stroke: potential value
of machine learning algorithms. Frontiers in neurology, 9, 784..

  The machine learning algorithms as Random Forests, Support Vector Machine, Neural Network, and Super Learner and
compared their predictive value with classic logistic regression models.​

• Random Forests is  used as an efficient screening tool to pick up predictive patterns in the dataset. It is be used as an
efficient screening tool to pick up predictive patterns in the data​.

• In the decision trees, each node represents a variable and splits the input data into branches based on an objective function
that determines the optimal threshold for separating the outcome classes.

• Logistic regression has the advantage of transparency at the level of individual variable coefficients, since from these
coefficients odds ratios can be derived.

• The outer CV(cross validation) loop with 100 repeated random splits, and obtained 100 different AUCs from every model.
We computed the average ROC-curve and mean AUC with 95% confidence intervals (CI) for all models.

• The random forest algorithm had highest accuracy score of 95.2% .the remaining algorithms the had performed with
accuracy ranges between 88% -95%.​
[8] Ho KC, Speier W, Zhang H, Scalzo F, El-Saden S, Arnold CW. A Machine Learning Approach for Classifying Ischemic
Stroke Onset Time From Imaging. IEEE Trans Med Imaging. 2019 Jul;38(7):1666-1676. doi: 10.1109/TMI.2019.2901445. Epub
2019 Feb 25. PMID: 30802855; PMCID: PMC6661120.

• A machine learning approach for TSS classification using routinely acquired imaging sequences and develop imaging
features from the magnetic resonance (MR) images and train machine learning models to classify TSS.

• Machine learning models have been applied widely and can achieve good classification performance for problems in the
healthcare domain because of their ability to learn and utilize patterns from data to make predictions.

• A deep learning model to extract hidden representations for the MR perfusion-weighted images and demonstrate
classification improvement by incorporating these additional deep features.

• The cross-validation results show that our best classifier achieved an area under the curve of 0.765, with a sensitivity of
0.788 and a negative predictive value of 0.609, outperforming existing methods.

• The features generated by our deep learning algorithm correlate with MR imaging features, and validate the robustness of
the model on imaging parameter variations.

• Determining stroke onset time independent of patient history is a challenging and important task for better stroke
evaluation and stroke treatment decision-making.
[9].M. Monteiro et al., "Using Machine Learning to Improve the Prediction of Functional Outcome in Ischemic Stroke
Patients," in IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 15, no. 6, pp. 1953-1959, 1 Nov.-Dec.
2018, doi: 10.1109/TCBB.2018.2811471.

• Stroke treatment is not risk-free, physicians only proceed with treatment when the potential benefits outweigh the perceived
risk.

• Metric for assessing the functional outcome of a stroke patient is the modified Rankin Scale (mRS). This 0 to 6 scale measures
the degree of disability or dependence of a patient as it relates to daily activities .

• The Acute Stroke Registry and Analysis of Lausanne (ASTRAL), the DRAGON  and the Totaled Health Risks in Vascular
Events (THRIVE) scores are meant to be calculated immediately when the patient is admitted in order to inform the treatment
decision.

• Machine learning techniques such as, regularized Logistic Regression, Decision Tree, Support Vector Machine (SVM), Random
Forest are used to predict the functional outcome of a patient.

• The Random Forest classifier gives higher accuracy of 93% and the other classifiers will gives an average accuracy of 84.3%.
[10]. R. Su et al., "autoTICI: Automatic Brain Tissue Reperfusion Scoring on 2D DSA Images of Acute Ischemic Stroke Patients,"
in IEEE Transactions on Medical Imaging, vol. 40, no. 9, pp. 2380-2391, Sept. 2021, doi: 10.1109/TMI.2021.3077113.

• The Thrombolysis in Cerebral Infarction (TICI) score is an important metric for reperfusion therapy assessment in acute ischemic
stroke.

• First, each digital subtraction angiography (DSA) acquisition is separated into four phases (non-contrast, arterial, parenchymal
and venous phase) using a multi-path convolutional neural network (CNN), which exploits spatiotemporal features.

• The MR CLEAN Registry  dataset for training and evaluating the proposed methods. The MR CLEAN Registry is an on-going
multi-center registry which contains all patients with acute ischemic stroke.

• The deep learning model for phase classification was developed using PyTorch on an NVIDIA 2080 Ti with 11 GB of memory
and The deep learning network was trained with a batch size of 32 for 100 epochs.

• The Weighted F1 score is used as evaluation metric for the Deep Learning model and it gives accuracy of 92%.
[11]. L. Tomasetti, L. J. Høllesli, K. Engan, K. D. Kurz, M. W. Kurz and M. Khanmohammadi, "Machine Learning Algorithms
Versus Thresholding to Segment Ischemic Regions in Patients With Acute Ischemic Stroke," in IEEE Journal of Biomedical and
Health Informatics, vol. 26, no. 2, pp. 660-672, Feb. 2022, doi: 10.1109/JBHI.2021.3097591.

• Ischemic stroke may be caused by arteriosclerosis, thrombi, emboli, dissections, or systemic hypoperfusion.All of them leading to
ischemia due to reduced blood flow in regions of the brain.

• According to the European Stroke Organization guidelines, Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) are
the two modalities recommended for diagnostic imaging in acute stroke patients.

• The necessary training set is available to train appropriate ML algorithms, good prediction models can be obtained. The ML models
tested in this study include Support Vector Machine, Decision Tree learning, and Random Forest.

• A preprocessing step is used to extract the brain tissue from the whole image and work with pixels within the brain tissue (BT), and
this step helps to balance the classes inside the dataset.

• Statistical measures to select the best input data combination to use. All the methods were tested with the best number of superpixel
regions.
​
[12]. V. L. Coli et al., "Detection of Brain Strokes Using Microwave Tomography," 2018 IEEE International Symposium on
Antennas and Propagation & USNC/URSI National Radio Science Meeting, 2018, pp. 223-224, doi:
10.1109/APUSNCURSINRSM.2018.860940 4.

• Ischemic stroke is due to a clot that blocks the blood circulation. An intravenous tissue plasminogen activator (tPA) must be
administered within 3 hours of stroke onset.

• The model of a synthetic brain obtained from CT and MRI data, which consists of a complex permittivity map of 362 × 434 × 362
data points with a spatial resolution of 0.5 mm, is used as a ground truth for the experiments.

• MRI can accurately detect both hemorrhagic and ischemic strokes, while CT can only detect hemorrhagic strokes but can be used to
determine if the stroke is ischemic through the process of elimination.

• The hemorrhage is detected at the right location in all tested cases, whereas no false detection occurs in case of ischemic strokes and
the binary image is obtained by pixel-wise multiplication and thresholding.

• Real and imaginary parts of the complex permittivity are computed separately. 

• Dielectric permittivity profiles were reconstructed separately on real and imaginary parts. The central cross-section of the head has
an elliptical shape and has higher intensity which gives accuracy of 75.3%.
[13] Pathanjali C , Monisha G , Priya T , Ruchita Sudarshan K, Samyuktha Bhaskar, 2020, Machine Learning for Predicting
Ischemic Stroke, INTERNATIONAL JOURNAL OF ENGINEERING RESEARCH & TECHNOLOGY (IJERT) Volume 09,
Issue 05 (May 2020)

• A small percentage of stroke patients die immediately from the initial trauma. Some of the leading causes to death may be initial
ischemic infarction, recurrent ischemic stroke, coronary artery disease  and other vascular or nonvascular causes.

• Cardiovascular Health Study (CHS)  dataset is very challenging to use effectively due to a significant fraction of missing
values and a large number of features in the dataset.

• If an artery is blocked, the brain cells (neurons) cannot make sufficient energy and will eventually stop working. If the artery
remains blocked for a couple of minutes, the brain cells may begin to die.

• Adaptive Neuro-Fuzzy Inference System (ANFIS) is composed of the artificial neural network in the first stage that inputs the
data and combines them in clusters based on their similarities. 

• These clusters are connected with the help of fuzzy rules to generate specific decision surfaces which calculate specific output.

• Support vector machine gives higher accuracy of 94.45% and Randon forest classifier will gives an accuracy of  93.44%.
[14].Li S, Zheng J, Li D. Precise segmentation of non-enhanced computed tomography in patients with ischemic stroke based on
multi-scale U-Net deep network model. Comput Methods Programs Biomed. 2021 Sep;208:106278. doi: 10.1016/j.cmpb.2021.106278.
Epub 2021 Jul 9. PMID: 34274610.

• The ischemic stroke was caused when the blood clotted in arteries, which gives blood supply to the brain. The image features of 30
stroke patients were segmented using a multi-scale U-Net deep network model.

• Computed tomography (CT) and magnetic resonance imaging (MRI) are essential tools for the assessment of acute ischemic stroke. In
the acute phase, CT and MRI both have high detection accuracy, sensitivity, and specificity. 

• U-Net is the generic network structure proposed to  combine low-level information with high-level information to improve the
segmentation effect. 

• In this paper U-Net model is used for image segmentation and to realize end-to-end automatic segmentation of ischemic stroke lesions.

• To Solve the class imbalance problem in the data we have used Dice's objective function and to enhanced the data for segmentation.
      These methods can achieve accurate pixel level positioning.

• Artificial Intelligence technologies require large databases to function effectively, imaging data must be collected systematically and
systematically.
[15].Yifan Hao, Aiyan Ji, Rongrong Xing, Wenqing Zhu, Baohong Jiang, Yi Jian, Hao Chen, Capillaries segmentation of NIR-
II images and its application in ischemic stroke, Computers in Biology and Medicine, Volume 147,2022,105742, ISSN 0010-
4825.

• A vessel segmentation algorithm was used to extract more reliable blood capillaries from a single NIR-II image and the algorithm
uses the mean of 8-connected neighbors to remove noise.

• A vascular segmentation algorithm  evaluate the quantified characteristics of main blood vessels and capillaries which were
essential for ischemic stroke evaluation.

• The algorithms detection accuracy and ability decrease with the increased depth, our algorithm has better performance than other
algorithms. 

• Functional magnetic resonance imaging (fMRI) and positron emission tomography (PET) can gain information about the brain’s
blood vessel and stroke area.

• The Dice similarity coefficient(DSC) and sensitivity(SN) are used as the evaluation metrics.

• DSC is between 0 and 1, with higher values indicating a better match between automatic segmentation map and ground truth.

• This method also can be used to estimate the degree of ischemic stroke and the algorithm has a specific guiding significance for
evaluating ischemic stroke and curative effect evaluation.
[16].Anjali Gautam, Balasubramanian Raman, Towards effective classification of brain hemorrhagic and ischemic stroke using
CNN, Biomedical Signal Processing and Control, Volume 63, 2021,102178, ISSN 1746-8094.

• The deep learning has been extensively used as a classification method because it automatically calculates features within the
convolutional layers of the deep system.

• Preprocessing algorithms have been used to remove some contrast abnormality so that better classification accuracy can be
achieved. In order to do this, we have made two copies of input images.

• The P_CNN model have been done by using stochastic gradient descent with momentum (SGDM) optimizer at time of training
the image dataset up to 30 epochs with 0.001 learning rate.

• The average classification accuracy obtained by AlexNet, ResNet50 and P_CNN_WP on the testing dataset is 87.77%, 91.11%
and 91.66% respectively.

• The proposed P_CNN model gives the best classification accuracy on three category dataset with an accuracy of 92.22% where
only 14 CT scan images have been wrongly identified. whereas for AlexNet, ResNet50 and P_CNN_WP their number counts to
22, 16 and 15 respectively.

• The total number of ischemic images are 350 and the number of ischemic images which are correctly identified are 256 which are
85.33% of total ischemic images, and 91% of normal images are correctly classified.
[17] To, M.N.N., Kim, H.J., Roh, H.G. et al. Deep regression neural networks for collateral imaging from dynamic susceptibility
contrast-enhanced magnetic resonance perfusion in acute ischemic stroke. Int J CARS 15, 151–162 (2020).
https://doi.org/10.1007/s11548-019-02060-7

• Magnetic resonance signals and the chosen parameters is automated data driven approach that  holds the potential for improving
the efficiency and effectiveness of collateral imaging.

• The dataset consists of 144 Acute ischemic stroke cases (AIS) and 201 Non-Acute ischemic stroke that includes the criteria and
symptoms of the ischemic Strock.

• Principal component analysis (PCA) were applied to extract blood vessel features, and SVM was used to predict the collateral
status achieving an accuracy of 82.2%

• In this paper the accuracy scores are measured using the squared correlation coefficient (R-squared), mean absolute error (MAE),
Tanimoto measure (TM), and structural similarity index (SSIM).

• The deep learning algorithms convolutional neural network had obtained the  highest accuracy scores as 0.897 R-squared, 0.946
TM, and 0.846 SSIM.
[18] Amish Kumar, Neha Upadhyay, Palash Ghosal, Tamal Chowdhury, Dipayan Das, Amritendu Mukherjee, Debashis Nandi,
CSNet: A new DeepNet framework for ischemic stroke lesion segmentation, Computer Methods and Programs in Biomedicine,
Volume 193,2020,105524,ISSN 0169-2607.

• In this paper the main purpose is deep learning model based on self -similar fractal networks and the U-Net model for performing
acute stroke diagnosis tasks automatically.

• The U-Net architecture is superseding other state of art architectures in the task of medical image segmentation. CNN architecture
consists of an encoder decoder network covers a contracting path capturing context and a balancing expanding path enabling
localization.

• They had incorporate the concept of a cascaded Classifier-Segmented architecture where the classifier selects only the slices which
contain lesions.

• In convolutional operation the SoftMax layer has been added at the output for the final prediction of the lesion and non-lesion
images.

• To improve the accuracy of the mode fold cross-validation is performed by dividing the dataset into five equal parts. Four parts
are used for training and one is for validation by this they obtained 94.78%.
[19] L. Zhang et al., "Ischemic Stroke Lesion Segmentation Using Multi-Plane Information Fusion," in IEEE Access, vol. 8, pp.
45715-45725, 2020, doi: 10.1109/ACCESS.2020.2977415.

• Ischemic stroke is the most common type of strokes it making up with 87 percent of  Cerebrovascular accident (CVA). It’s caused by
a blockage in an artery supplying the brain with blood. Hemorrhagic strokes which account for the remaining 13 percent.
• In this data is augmentation and the Nonnegative matrix factorization (NMF) method optimizes segmentation results through 3
Dimensional information and shape prior. 3-D fully convolutional densenets (FCD) extracts more effective features from data due to
its complex structure.
• This paper  focus on the ischemic stroke lesion segmentation. A fully automated method based on feature pyramidal networks
(FPN) based architecture as the detection network for extracting the patches from the entire image and utilize the U-NET .
• To reduce the information loss caused by the Region proposal network (RPN) thee candidate regions is dilated for the feature
extraction module modifies the multi-scale feature maps of FPN.
• The data Is preprocessed in three levels they are Data with N4bias correction, Mapping to the UINT8 data type directly and Three
window levels and window widths concatenate as to normalize, reduce bias and noises from the data.
• The detection and segmentation network (DSN) model had achieved the highest accuracy of 98.36%  where RCNN is achieved 98.1
accuracy with F-score as 21.64.
[20] Bin Zhao, Zhiyang Liu, Guohua Liu, Chen Cao, Song Jin, Hong Wu, Shuxue Ding, "Deep Learning-Based Acute Ischemic
Stroke Lesion Segmentation Method on Multimodal MR Images Using a Few Fully Labeled Subjects", Computational and
Mathematical Methods in Medicine, vol. 2021, Article ID 3628179, 13 pages, 2021. https://doi.org/10.1155/2021/3628179
• stroke can be divided into ischemia and hemorrhage based on the types of cerebrovascular accidents, where ischemic stroke accounts for
87%multimodal resonance image (MRI) are used to identify the stroke.

• A convolutional neural network (CNN) had presented an exceptional performance in computer vision. By training on a large number of
fully labeled data .CNN based method had shown great potentials in segmenting ischemic stroke lesion on the MRIs.
• The dice coefficient (DC) is used to evaluate the pixel-level segmentation performance. 

• It measures the overlap between the predicted segmentation P and the ground truth G. A multifeatured map fusion network (MFMF-
Network) is proposed and trained on the weakly labeled subjects to extract features .

• The data set consists of 582 subjects with Autery ischemic Strock (AIS) lesions  collected from retrospective database of Tianjin Huanhu
Hospital .s 398 weakly labeled subjects and 184 fully labeled

• This model had attained the accuracy scores as high precision rate of 0.852 and recall rate of 0.923.Mean dice coefficient of 0:651 ±
0:183 from the pixel level metric
[21] Samak, Z. A., Clatworthy, P., & Mirmehdi, M. (2022). FeMA: Feature matching auto-encoder for predicting ischaemic
stroke evolution and treatment outcome. Computerized Medical Imaging and Graphics, 99, 102089.

•  A novel framework is proposed in this paper to estimate the one week follow-up scan and MRS score as ischemic stroke treatment
outcome.

• FeMA consists of  two stages, predicting ischemic stroke evolution at one week or 90days.

• One week follow-up is without voxel-wise annotation and 90 days from a baseline scan.

• Feature similarity and consistency objective are used to get good follow-up scan representation from base line scan.

• In MRS scan U-Net , non U-Net models and a Feature Matching Auto-encoder (FeMA) is used. 

• Accuracy, F1-score, AUC are used as metrics in this paper.

[22] Zhao, B., Liu, Z., Liu, G., Cao, C., Jin, S., Wu, H., & Ding, S. (2021). Deep learning-based acute ischemic stroke lesion
segmentation method on multimodal MR images using a few fully labeled subjects. Computational and Mathematical Methods in
Medicine, 2021.

• A deep learning-based method  is implemented in this  paper using a few fully labelled subjects for AIS lesion segmentation.

• The proposed method consists of three processes: classification, segmentation, and inference.

• In the classification process, the network is trained on the weakly labelled subjects as a classifier to obtain a set of pretrained
parameters.

•  In the segmentation process, the network freezes the pretrained parameter and is further trained on the fully labelled subjects.

• CNN-based methods are used to show their great potentials in segmenting ischemic stroke lesions on the MRIs.

• By training on a large number of fully labelled subjects where the stroke lesions were annotated in a pixel-by-pixel manner.

• Dice coefficient and precision is used as performance metrics.

  
[23] Platscher, M., Zopes, J., & Federau, C. (2022). Image translation for medical image generation: Ischemic stroke
lesion segmentation. Biomedical Signal Processing and Control, 72, 103283.

• In this paper a Deep Learning based data augmentation technique is used to improve performance.

• Depending on the available clinical samples, the improvement can be substantial.

• Different image-to-image translation models  are used to synthesize magnetic resonance images of brain volumes with and without stroke lesions from
semantic segmentation maps.

• The two consecutive generative models are used , one for generating realistic stroke lesion labels and another to translate brain segmentation masks
into DWIs.

• The  ITMs  are used as 2D CNNs in tensor flow  are implemented and trained them on a single GPU for 100 epochs on the combined clinical normal
and stroke databases.

• The performance of the various models is evaluated using a U-Net which is trained to segment stroke lesions on a clinical test set.

• The Dice score of 72.8% for the model with the best performance, which outperforms the model trained on the clinical images alone 67.3% and is
close to the human inter-reader Dice score of 76.9%. 
[24] Islam, M., Vaidyanathan, N. R., Jose, V., & Ren, H. (2018, September). Ischemic stroke lesion segmentation using
adversarial learning. In International MICCAI Brainlesion Workshop (pp. 292-300). Springer, Cham.

• An automatic ischemic stroke lesion segmentation is used in this paper using adversarial learning scheme.

• To improve the performance of the segmentation architecture we demonstrated adversarial learning like U-Net.

• In this paper the proposed model consists of two deep networks and which are work at a time.

• The two models are classified as Generative model and Discriminative model.

• Generative model generates the synthesized model and discriminative model estimates the probability that a output came from
original data other than generative model.

• Precision, Recall are used as performance metric for this model.

[25] Kappelhof, N., Ramos, L. A., Kappelhof, M., van Os, H. J., Chalos, V., van Kranendonk, K. R., ... & Marquering, H. A.
(2021). Evolutionary algorithms and decision trees for predicting poor outcome after endovascular treatment for acute ischemic
stroke. Computers in Biology and Medicine, 133, 104414.

• In this paper a novel evolutionary algorithm for fuzzy decision trees are used to identify the patients with poor outcome after the
treatment.

• A modified Rankin Scale score is used  to define the endovascular treatment.

•  Explorative and stochastic techniques are used to perform larger and non greedy search for different decision trees.

• The fuzzy decision trees can  overcome the convolutional CART algorithm

• The trees which are created by evolutionary algorithm is comparatively smaller than CART algorithm.

• F1 score is used as a performance metric for this model.

[26] Zhang, S., Wang, J., Pei, L., Liu, K., Gao, Y., Fang, H., ... & Xu, Y. (2022). Interpretable CNN for ischemic stroke subtype
classification with active model adaptation. BMC medical informatics and decision making, 22(1), 1-12.

• This paper mainly focuses on the classification of Trial of Org 10172 in Acute Stroke Treatment (TOAST) subtype. The authors
proposed a novel deep learning architecture to address the problem.

• The dataset used in this paper was collected from the patients admitted to the dept. of neurology from 2014 to 2016 in AAA hospital.
It consists of a totally of 2310 patient data.

• XGBoost algorithm was used in data pre-processing to rank the original features. The features which were having feature
importance <= 0.005 are dropped and the remaining 93 features were taken by the authors.

• The entire framework consists of three components: Casual CNN architecture, Active selection criterion and KL-focal Loss.

• The authors proposed model achieved 60.2% accuracy, 0.6757 AUC score, 0.602 Recall and 0.6213 precision and they compared
their results which other architectures, pre-processing techniques and different loss functions to their model.
[27] Shi, T., Jiang, H., & Zheng, B. (2021). C 2 MA-Net: Cross-Modal Cross-Attention Network for Acute Ischemic Stroke
Lesion Segmentation Based on CT Perfusion Scans. IEEE Transactions on Biomedical Engineering, 69(1), 108-118.

• This paper mainly focuses on Acute ischemic stroke lesion segmentation based on CT perfusion scans. The authors demonstrated the
advantages of C2MA-network to segment the AIS lesions.

• The authors proposed a network which used a C2MA module directly to establish a spatial-wise relationship by using the multigroup
non-local attention operation between two modal features and performs dynamic group-wise recalibration through group attention
block.

• ISLES-2018 dataset is used to build and evaluate the model. The dataset consists of CT and CTP images of 103 patients with AIS
lesions.

• Dice Similarity Coefficient (DSC) is used as the main metric to evaluate the performance of the model. The C 2MA-Net achieved a
final mean DSC of 0.55 on training data and 0.48 on testing data.
[28]. H. Kuang, B. K. Menon and W. Qiu, "Segmenting Hemorrhagic and Ischemic Infarct Simultaneously From Follow-Up
Non-Contrast CT Images in Patients With Acute Ischemic Stroke," in IEEE Access, vol. 7, pp. 39842-39851, 2019, doi:
10.1109/ACCESS.2019.2906605. 

• In this paper, the post-treatment of Cerebral infarct volume (CIV) has measured by using the method makes use of advances in deep
learning and convex optimization techniques like convolution neural network learned sematic information, local image context. 

• The novel semi-automated segmentation is implemented to segment the ischemic and hemorrhage infarct. 

• By using the dice similarity coefficient and the mean and maximum absolute surface distance ,30 patient images are segmented. 

• CNN, D-U Net and the random forest are used to get the results.  The performance metrices like DSC,MAD,MAXD are used to
classify the performance of the models. 
[29] Clèrigues, A., Valverde, S., Bernal, J., Freixenet, J., Oliver, A., & Lladó, X. (2019). Acute ischemic stroke lesion core
segmentation in CT perfusion images using fully convolutional neural networks. Computers in biology and medicine, 115, 103487.

• In this paper, they presented and evaluated an automated method for acute stroke lesion core segmentation from CT and CTP images. 

• The presented tool achieves state-of-the-art performance while using a simple training procedure with a single network. 

• The dataset with 94 training and 69 testing patient sample are used for evaluating the ischemic stroke lesion segmentation. 

• Several metrices like Dice similarity coefficient, MAD are used to calculate the performance of the model. 
[30] Zhao, B., Ding, S., Wu, H., Liu, G., Cao, C.,Jin, S., & Liu, Z. (2019). Automatic acute ischemic stroke lesion segmentation
using semi-supervised learning. arXiv preprint arXiv:1908.03735. 

• This paper mainly focus to localize and quantitively evaluate the acute ischemic stroke (AIS) lesions. 

• The CNN models are trained on hundreds of fully-labelled subjects with accurate annotations of ASI lesion. 

• A double-path classification network (DPC-Net) was proposed and trained using the weakly-labelled subjects to detect the suspicious
AIS lesions. 

• By using 460 weakly-labelled subjects and 5 fully-labelled subjects to train and fine-tune the proposed method and achieves a mean
dice coefficient of 0.642, and a lesion-wise F1 score of 0.822 on a clinical dataset with 150 subjects. 
 COMPARISON TABLE
SNO Title Year   Description Limitations Advantages Performance Gaps
Metrics

A novel automatic method to Important issue of 3D CNN gives

Dice similarity
Zhang, R., Zhao, L., segment acute ischemic stroke automatic acute stroke accurate and robust Investigating how to borrow
coefficient,
Lou, W., Abrigo, J. from diffusion weighted images segmentation from DWI performance,and it is power from other medical or
images is to accurately good potential in lesionwise precisio natural data, e.g. by transfer
1 M., Mok, V. C., Chu, 2018 (DWIs) using deep 3-
and lesionwise F1
W. C., ... & Shi, D convolutional neural networks reject the numerous clinical practices, and learning, to reduce our
artifacts present on DWI can serve as a score model’s dependency on large
L. [1] (CNNs) is used.
scans while retaining true preliminary step in a annotated datasets .
stroke lesions CAD.

 CNNs use a
A convolutional neural network In the future, we will collect
variation of multilayer
(CNN) is a class of deep , feed- more brain stroke images,
CNN recognition rate will  perceptrons designed 
2 forward artificial neural network TPR, FNR, FPR, and and input the image into
C. -L. Chin et al. [2] 2017 be higher and runtime is  to require
that have successfully TNR CNN for training to enhance
more. minimal preprocessin
been applied to analyzing visual the overall system
g.
imagery. recognition rate

Demonstrating a machine
Magnetic resonance Sensitivity,
learning approach for (Time
K. C. Ho, W. Speier, The machine learningimaging (MRI) specificity, F1-score,Collecting the Computed
Series Stroke) TSS classification
3 H. Zhang, F. Scalzo, 2019 models were trained andanalysis one step positive predictiveTomographic (CT) perfusion
using routinely acquired
S. El-Saden and C. validated on only the MRcloser to an value (PPV), andimages and validating the
imaging sequences and
W. Arnold. [3] images. operational decision negative predictiverobustness.
correlation method is also used.
support. value (NPV) .
SNO Title Year   Description Limitations Advantages Performance Gaps
Metrics

A predictive model
was introduced for seizure
occurrence in a single patient. Advanced imaging
This approach was based on a modalities in
4 Kamal H, Lopez V, 2018 multi-class support vectorIt takes more time duringidentifying patients A biomarker of cerebral
Sheth SA. [4] machine (SVM) and 14 selectedthe training the model. with a higher forAccuracy, Precision edema severity.
features of an better outcomes from
electroencephalogram in EST
patients with epilepsy.

Support vector machine (SVM)

using n-
gram technique obtained lesser The differences in
F1 score so to improveGeneralizable resultsclass proportion in the The model advancement
5 Kim C, Zhu V, Obeid the measure applied bag of wordcould occur if we usetraining dataset willF1 score
2019  with the deep learning model
J, Lenert L. [5] technique to add additional
features. those NLP and ML tools. not affect the and get better accuracy.
outcome.

A deep supervised approach for Compared to other

Karthik, R., Gupta, U., It can be extended
6 ischemic lesion segmentation Damaged tissues are not works in this area this
Jha, A., Rajalakshmi, 2019 Dice coefficient for delineation of core and
from multimodal MRI using properly segmented. work showed better
R., & Menaka, R.  penumbra region separately
Fully Convolutional Network results
SNO Title Year   Description Limitations Advantages Performance Metrics Gaps

Van Os, H. J., Ramos,

Predicting Outcome of Best performing logistic Predictive patterns in the
L. A., Hilbert, A., Van
Endovascular Treatment for AcuteDue to their greatregression models inConfidence intervals data may be missed, as
7 Leeuwen, M., 2018 Ischemic Stroke: Potential Value offlexibility machinepredicting radiological or variable selection is strictly
Van Walderveen, M.
Machine Learning Algorithms learning algorithm areclinical outcome in stroke based on the literature and
A., Kruyt, N. D., ... &
prone to overfitting. patients. expert opinion
Mr Clean. [7]

A deep learning model to extract

hidden representations for the MR Sensitivity, specificity,
Magnetic resonance
Ho KC, Speier W, perfusion-weighted images andThe machine learningimaging (MRI) analysisF1-score, positiveCollecting the Computed
8 Zhang H, Scalzo F, El- 2019 demonstrate classification models were trained and predictive valueTomographic (CT)
one step closer to an
Saden S, Arnold CW. improvement by incorporatingvalidated on only the MRoperational (PPV), and negativeperfusion images and
decision
[8] these additional deep features. images. predictive valuevalidating the robustness.
support.
(NPV) .

Machine learning techniques such The machine learning

aim to incorporate the use
as, regularized Logistic Regression, Training the classifiers techniques can be
of image and genetic
Decision Tree, Support Vector with more data could meaneffectively used to predict
9 2018 AUC and ROC curves information and to take
M. Monteiro. [9] Machine (SVM), an increase in the average the functional outcome of
and Rankin Scale. advantage of the
Random Forest are used to predict AUC, and difficult in an ischemic stroke patient
longitudinal aspect of the
the functional outcome of a patient. calculation of scores. three months after the
data.
initial event.
SNO Title Year   Description Limitations Advantages Performance Gaps
Metrics

The Thrombolysis in Cerebral In the MR CLEAN Both phase

R.Su et al., "autoTICI: Infarction (TICI) score is an Registry, one of the classification and
Automatic Brain Tissue important metric for limitations is that the perfusion F1 Score In future we implement a AI
10 Reperfusion Scoring on 2021 reperfusion therapy assessment original un subtracted segmentation can Model to identify the stroke
2D DSA Images of Acute in acute ischemic stroke. images were mostly benefit from deep part.
Ischemic Stroke Patients," discarded, leaving only learning methods in
the subtracted ones. this work.

A preprocessing step is used to

Core predictions are
extract the brain tissue from the The Two-Step In the future, we plan to use
slightly slower than the
L. Tomasetti, L. whole image and work with approach with RF approaches based on deep
one presented by the
11 J. Høllesli , K. Engan, K. 2022 pixels within the brain tissue
thresholding methods
produces slightly Accuracy, Dice neural networks with 4D
D. Kurz, M. W. Kurz and (BT), and this step helps to better results in the Coeff Computed Tomography
regardless of the group,
Khanmohamma.[11] balance the classes inside the Hausdorff distance Perfusion (CTP) volume as
while penumbra
dataset. for the core class. input.
predictions are superior.

Detection of Brain Strokes Identification,

V. L. Coli et al., Using Microwave Tomography. including image
In the future we will
12 "Detection of Brain 2018 The model of a synthetic brain reconstruction and
investigate specific processing
Strokes Using Microwave obtained from CT and MRI automatic detection,
for the ischemic stroke.
Tomography,"  data. takes less than five
minutes.
SNO Title Year   Description Limitations Advantages Performance Gaps
Metrics

Adaptive Neuro-
Fuzzy Inference System
Pathanjali C , The model removes
(ANFIS) is composed of the
Monisha G , Different machinethe redundancy of Optimal feature sets for
artificial neural network in the
13 Priya T , Ruchita    2020 learning methods mayproviding the same each machine learning
first stage that inputs the data   confusion matrix
Sudarshan K, not perform equally ondatasets over and methods were defined
and combines them in
Samyuktha the same feature set. over again for systematically.
clusters based on their
Bhaskar prediction. 
similarities. 

Li S, Zheng J, Li Computed tomography (CT) A multi-scale U-

D. Precise and magnetic resonance Net architecture
Collaborative efforts across
segmentation of imaging (MRI) are essential The cross entropy loss model was proposed
multiple hospitals and
14 non-enhanced     2021 tools for the assessment of training model converges to target
   Loss Function centers are essential and
computed acute ischemic stroke. In the slowly, and the training cerebral infarction
will be an important
tomography in acute phase, CT and MRI both curve is not stable. lesions in on-
direction of future work.
patients with have high detection accuracy, enhanced CT images
ischemic stroke. sensitivity, and specificity accurately.

Yifan Hao, Aiyan

A vascular segmentation
Ji, Rongrong Xin In evaluating
algorithm  evaluate the  It is difficult to eliminate
g, Wenqing ischemic stroke, the In the next stage, we will
Zhu, Baohong Jia     2022
quantified characteristics of the influence of breathing
vascular quantization
  Dice Similarity research the related
15 main blood vessels and and heartbeat, especially
ng, Yi Jian, Hao algorithm based on Coefficient, registration methods of
capillaries which when the integration time Sensitivity.
Chen, Capillaries our algorithm can respiratory and heartbeat. 
were essential for ischemic is long. 
segmentation effectively
stroke evaluation.
of NIR-II images.
 Performance Met
SNO​ Title Year  ​ Description​ Limitations​ Advantages​ Gaps​
rics​

Anjali Gautam,
The deep learning has been ​ he thresholding
T
Balasubramanian ​ he size of P_CNN is
T
extensively used as a operation
Raman, Towards very small it gives better
classification method because  performed for ​ recision,F-
P I​ n future we are
effective classification accuracy
it automatically calculates reducing the size of measure,TPR,FPR implementing multifeature
16 classification of 2021 as compared to
features within feature set by extraction using simple
brain AlexNet and ResNet50
the convolutional layers of the excluding all ​ deep learning techniques.
hemorrhagic and in very less execution
deep system. unnecessary pixel
ischemic stroke time.
​ values.
using CNN.
​ ean
M R-
​ eep
D regression
​ inh
M Nguyen square, mean
neural networks for ​ ive
F collateral
Nhat To· Hyun ​ ess accurate prediction
L absolute error
collateral imaging from phase mapping and
Jeong Kim · Hong of the 3D-DRNN for (MAE), Tanimoto ​ he evaluation of the
T
17 ​2019 dynamic susceptibility deep neural
Gee Roh · Yoon- both arterial and late measure (TM) and robustness to the artifacts. 
contrast-enhanced network for
Sik Cho ·Jin Tae venous phase. structural
magnetic resonance perfusion collateral images.
Kwak similarity index
in acute ischemic stroke.
(SSIM).
​ mish
A Kumar,
​ NN
C based
Neha Upadhyaya ,
I​ ncorporating architecture covering a
Palash Ghosal, ​ new DeepNet framework
A
I​ t take 8 hours of time to denser feature ​ ice coefficient,
D Greater number of classes
18 Tamal ​2020 for ischemic stroke lesion
train the model. extraction layers at recall, precision. to detect stroke lesions
Chowdhury, segmentation.
the encoder side. based on acute CT
Dipayan Das,
perfusion data.
Amritendu M
SNO Title Year   Description Limitations Advantages Performance Metrics Gaps

LONG ZHANG , A triple-branch DSN architecture, The labeled data is scarce, Dice Highlight areas in a given
2D segmentation
19 RUONING  S,2020 used for extracting different planewhich may lead to coefficient, accuracy, input image that provide
frameworks such as U-
YUANYUAN WANG, feature, multi-plane fusionoverfitting and the label with sensitivity, specificity, useful evidence for lesion
Net, Deep Lab.
CHUANG ZHU , JUN LIU network (MPFN). insufficient accuracy. recall and F-Measure. segmentation.

Deep Learning-Based Acute Greatly reduce the

Bin Zhao ,  Zhiyang Liu , The model had detection rate
Ischemic Stroke Lesion expense of obtaining aMean diceTo include the pretrained
20  Guohua Liu, Chen Cao,2021 (DR) to measure missed
Segmentation Method on large number of fullycoefficient, precision ratemodels for training the
Song Jin,2 Hong Wu, and subjects as a subject-wise
Multimodal MR Images Using a labeled subjects in aand recall rate. model.
Shuxue Din metric.
Few Fully. supervised setting.

The proposed model Feature  ​The feature similarity

Matching Auto encoder is used and consistency
​It gives good accuracy but Accuracy, F1-score,
Samak, Z. A., Clatworthy, P., to illustrate the process for supervision to obtain ​ isunderstanding important
M
21 2022 the main problem is too AUC 
& Mirmehdi, M. [21] predicting ischemic stroke better follow-up scan variables.
lossy. ​
evolution, as well as mRS scores, representation from the
given a baseline scan on hospital baseline scan. 
admission.
 Performance 
SNO​ Title Year  ​ Description​ Limitations​ Advantages​ Gaps​
Metrics​
A convolutional neural network
(CNN) had presented an exceptional I​ n the process of
​ hao, B., Liu,
Z performance in computer vision. By I​ t gives the high
feature extraction the ​ ixel quality is reduced
P
Z., Liu, G., Cao, training on a large number of fully performance because of ​ ice
D
labeled data .CNN based method pixel size will reduce after  feature extraction, then
22 C., Jin, S., Wu, ​2021 CNN model, CNN can coefficient,
had shown great potentials in at the same time automatically it will lose
H., & Ding, S.  handlel larger data sets Precision
segmenting ischemic stroke lesion image quality also some information.
[22] very easily.
on the MRIs. reduces. 

modified network
Different image-to-image architecture could be
translation models  are used to constructed, such as the
​ latscher,
P synthesize magnetic resonance ​ raining on synthetic
T
one by which allows ​ raining on synthetic data
T
M., Zopes, J., images of brain volumes with and data alone does not a​ ccuracy,
23 ​2022 users to easily adapt the alone does not provide
& Federau, C. without stroke lesions from provide competitive precision
models to their data, competitive results.
[23] semantic segmentation maps. results.
even if acquired under
entirely different
circumstances. 

 ​Because of poor
An automatic ischemic stroke lesion visibility of the lesion The proposed  model ​ he overall
T
Islam, M., segmentation is proposed using area in consists of two deep Precision, performance of the model is
Vaidyanathan, N. adversarial learning scheme. An the CT imaging, the networks and which are Recall  not excellent because of poor
24 2018
R., Jose, V., & adversarial learning can improve the overall performance of work at a time, that visibility of the lesion area in
Ren, H. [24] performance of the segmentation    the model means time will reduce. ​ the CT imaging.
architecture like U-Net. is not excellent.
 Performance 
SNO​ Title Year  ​ Description​ Limitations​ Advantages​ Gaps​
Metrics​

​ his stochastic nature

T
​ appelhof, N.,
K
The  proposed model can prevent the
Ramos, L.
implemented all ITMs as 2D algorithm from getting
A., Kappelhof, I​ t does not provide the
CNNs in tensor flow and I​ t take more iterations stuck in a local
M., van Os, H. J., necessary subtlety we
trained them on a single GPU because of remove the optimum, improving
25 Chalos, V., van ​2021  accuracy desire: insignificant splits
for 100 epochs on the selection pressure, it chances of finding the
Kranendonk, K. on small sets of data can
combined clinical normal and will take more time. global optimal
R., ... & still occur. 
stroke databases. solution where
Marquering, H. A.
deterministic
[25]
algorithms. 

The authors did not

 ​A deep learning framework is considered any feature
​ hang, S., Wang,
Z An efficient algorithm
proposed  which can be combinations which
J., Pei, L., Liu, K., which uses less This paper did not focus on
26 ​2022 efficiently used in the Acute might increase the Precision,
Gao, Y., Fang, computation to the subtype classification
Ischemic stroke subtype performance of the recall, F1-Score
H., ... & Xu, Y classify the subtype of using stroke images.
classification. model.
AIS.

This paper mainly focuses on Compared to other

Acute ischemic stroke lesion C2MA network is used
Shi, T., Jiang, H., segmentation based on CT works, in some cases DSC, precision,
27 2022 for the segmentation
& Zheng, B. perfusion scans using C2MA- this work is not efficient recall, F1-score
of the AIS Lesions
Net in performance
 Performance 
SNO​ Ti Year  ​ Description​ Limitations​ Advantages​ Gaps​
Metrics​

HULIN KUANG ,
BIJOY K.
Machine learning
MENON, AND
The mean DSC of based methods, semi Dice similarity
WU QIU
In the novel semi-automated 67.4% obtained by the RF, semi-U-Net, and coefficient,
Department of U-Net and D-Unet, and
segmentation is implemented to proposed method is Semi-D-Unet, mean and
28 Clinical 2019 significantly decreased the
segment the ischemic and higher than 58-62%  but performed better than maximum
Neurosciences, MAD by around 8 times.
hemorrhage infarct.  lower than a median of the classic graph cut absolute
University of
74% (42-90%)  method relying on surface.
Calgary, Calgary,
image context
AB T2N 4N1,
Canada

Albert Clèrigues  ,
In this paper, they presented and
Sergi Valverde, Prediction dropout It does not explain about
evaluated an automated method for The training requires
Jose Bernal, Jordi layers are used to DSC,PPV,HD, the deep layers in the fully
29 2019 acute stroke lesion core minimal tuning of
Freixenet, Arnau reduce the outliers R2, Sensitivity. convolutional  neural
segmentation from CT and CTP parameters 
Oliver, Xavier networks.
images. 
Lladó

In k-means
Bin Zhaoa ,  Clinically, the LI stroke
clustering the clustering it presents very high
Shuxue Dinga,b , accounts for 85% of all
This paper mainly focus to localize lesion regions will be precision of 0.880,
Hong Wua , Dice coefficient AIS patients. However, it is
larger than the true which is of paramount
30 Guohua Liua , 2019 and quantitively evaluate the acute ,f1 score, much difficult to be
lesions. As the K importance in
Chen Caoc , Song precision. diagnosed in clinical
ischemic stroke (AIS) lesions increases, the clustering avoiding misdiagnosis
Jinc , Zhiyang practice, especially when it
lesion regions will in clinical scenario.
Liua, is too small to be noticed. 
gradually decrease.
                  METHODOLOGY
Dataset overview:  For the classification of type of  ischemic stroke (CE or
LAA) the  Mayo Clinic Strip-Ai dataset was considered. It is having more than
700 images with labels as CE or LAA. Each image consists of high resolution in
the range of 100MB to 2GB which are stored in the format of Tag Image File
Format(TIFF).
             In order to train the models with this large size data, will be a tedious task
to get  better accuracy. Such that there are several alternative models to decrease
the resolution of the images without losing the content or features.

Removing the unwanted pixels form the image:  It is observed that the images
are having unwanted white spaces which will increase the size of image. Pruning
is applying on every image where there will a threshold value such that if the
pixel meets the threshold value the pixel will get terminated. Fig 1.original image Fig 2.image after
After applying the pruning process for every image. The image size will reduced removing the unwanted
white space
to the one third of the real image as shown  in the fig. Every pruned or
compressed image is saved in other folder for the use of further process.
                  METHODOLOGY
Resizing the images: Now the images in the dataset contains
different shapes. For example one of CE's label image is having
the shape (10816,7645,3) and other one is
having(20345,8465,3). In this constraint, if the images are
applied to any model without any fixed shape(1024,512,3), the
model throws an error. So, there should be a fixed shape for
every image.  Image resize is one of the methods where the
image will be resized in to desired shape. For retrieving
features from the images ,the HOG model is used. There is a
constraint that every input image for the HOG model should be
in the format where the height of image should twice the width
of image.
Fig2 compressed imager
Fig3. Resizing the image into
fixed size
(1024,512,3)
                  METHODOLOGY
Data Augmentation: It is one of the techniques to balance the imbalanced dataset. The mayo clinic Strip- AI
dataset contains 500 CE labels and 200+ LAA labels. In order to balance the dataset there are two ways i.e., data
under sampling and data over sampling.
   For the dataset, data over sampling Technique is used to increase the minority class labels. After Augmentation
the dataset will be reaches to 1000+ images with balanced data.
The oversampling can be achieved by several ways:
1.Random Over Sampling
2.Smote
3.KMeans Smote
   The Smote oversampling Augmentation technique is used to balance the dataset. Smote is the acronym of
Synthetic Minority Oversampling Technique. 
The steps involved in Smote are:
Step1:Select a sample, let’s call it O(for Origin), from the minority class randomly
Step2:Find the K-Nearest Neighbours of O that belong to the same class
Step3:Connect O to each of these Neighbours using a straight line
Step4:Select a scaling factor ‘z’ in the range [0,1] randomly
Step5:For each new connection, place a new point on the line (z*100)% away from O. These will be our
synthetic samples.
Step6:Repeat this process until you get the desired number of synthetic samples.
HOG Model :
• Histogram of Oriented Gradients, also known as HOG, is a feature descriptor like the Canny Edge Detector, SIFT
(Scale Invariant and Feature Transform) . It is used in computer vision and image processing for the purpose of
object detection. The technique counts occurrences of gradient orientation in the localized portion of an image.  
• The HOG descriptor focuses on the structure or the shape of an object. It is better than any edge descriptor as it
uses magnitude as well as angle of the gradient to compute the features. For the regions of the image it generates
histograms using the magnitude and orientations of the gradient.

Steps to calculate HOG features:

1. consider the images in the shape where the height of image is twice the width of the image. In the  Mayo
Clinic Strip-Ai dataset the images are in random sizes. After resizing the images into  the format of 1024X512 the
dataset will be applied to the HOG model.
2.After resizing the images the Hog will calculate the Gradient of images. The gradient is obtained by combining
magnitude and angle from the image. Considering a block of 3x3 pixels, first Gx and Gy is calculated for each pixel.
First Gx and Gy is calculated using the formulae below for each pixel value .
After calculating Gx and , magnitude and angle of each pixel is calculated using the formulae mentioned below.
             Magnitude(µ)=√(Gx­² +Gy²)
              Angle(Θ)=|tan{}^{-1(Gy/Gx)|
3. After obtaining the gradient of each pixel, the gradient matrices (magnitude and angle matrix) are divided
into mXm cells to form a block. For each block, a n-point histogram is calculated. A n-point histogram develops a
histogram with n bins and each bin has an angle range of  x degrees.
4. Once histogram computation is over for all blocks, 4 blocks from the n point histogram matrix are clubbed
together to form a new block (m X m). This clubbing is done in an overlapping manner with a stride of 8 pixels.
For all 4 cells in a block, we concatenate all the n point histograms for each constituent cell to form a X feature
vector.
5. Normalization is done to reduce the effect of changes in contrast between images of the same object. From
each block a X point feature vector is collected. In the horizontal direction there are several blocks and in the
vertical direction there are several blocks. So the total length of HOG features will be : horizontal X vertical X x=
Feature descriptor. HOG features of the selected image are obtained.

                                     The resize image into Hog Gray scaled Image                                                                   

Decision Tree:
       Decision Tree can be used for both classification and Regression problems, but mostly preferred for 
Classification problems. It is a tree-structured classifier, where internal nodes represent the features of a dataset,
branches represent the decision rules and each leaf node represents the outcome.
STEPS:
     Step-1: Begin the tree with the root node, says S, which contains the complete dataset.
     Step-2: Find the best attribute in the dataset using Attribute Selection Measure (ASM).
     Step-3: Divide the S into subsets that contains possible values for the best attributes.
     Step-4: Generate the decision tree node, which contains the best attribute.
     Step-5: Recursively make new decision trees using the subsets of the dataset created in step-3.
Continue this process until a stage is reached where you cannot further classify the nodes and called the final node as
a leaf node.
In Decision Tree the major challenge is to identification of the attribute for the root node in each level. This process is
known as attribute selection. We have two popular attribute selection measures:
1. Information Gain: Finding a node in a decision tree to partition the training instances into smaller subset
Information gain is measured. Information gain means change in entropy. Entropy is the measure of uncertainty of a
random variable, it characterizes the impurity of an arbitrary collection of         examples. The higher the entropy
more the information content.
2. Gini Index: It is a metric to measure how often a randomly chosen element would be incorrectly identified. It
means an attribute with lower Gini index should be preferred.
Random Forest Algorithm :
 Random Forest is a classifier that contains a number of decision trees on various subsets of the given dataset and
takes the average to improve the predictive accuracy of that dataset. The Random Forest Classifier  shows many
different parameters, we can select for our model. Some of the important.

1. n_estimators — the number of decision trees you will be running in the model.
2. max_depth — this sets the maximum possible depth of each tree.
3. max_features — the maximum number of features the model will consider when determining a split.
4. bootstrapping — the default value for this is True, meaning the model follows bootstrapping
principles(defined earlier).
5. max_samples — This parameter assumes bootstrapping is set to True, if not, this parameter doesn’t
apply. In the case of True, this value sets the largest size of each sample for each tree.
6. Other important parameters are criterion, min_samples_split, min_samples_leaf, class_weights,
n_jobs,  and  others that can be read in the sklearn’s Random ForestClassifier documentation. 
 It takes less training time compared to other algorithms and it maintains accuracy when a large proportion of data
is  missing. It is capable of handling large datasets with high dimensionality.
XGBoost Algorithm:
 XGBoost is an implementation of Gradient Boosted decision trees. In this algorithm, decision trees are created in
sequential form. Weights play an important role in XGBoost. Weights are assigned to all the independent variables
which are then fed into the decision tree which predicts results.

 The weight of variables predicted wrong by the tree is increased and these variables are then fed to the second
decision tree. These individual classifiers/predictors then ensemble to give a strong and more precise model. It can
work on regression, classification, ranking, and user-defined prediction problems.

 XGBoost is an extension to gradient boosted decision trees (GBM) and specially designed to improve speed and
performance.

XGBoost Features
1. Regularized Learning
2. Gradient Tree Boosting
3. Shrinkage and Column Subsampling
 RESULTS AND DISCUSSIONS

VGG-16 EfficientNetB7
Dataset(Imbalanced) 67.54% 67.58%

Decision tree Random Forest XGBoost XGBoost(RF)

HOG features(Imbalanced Dataset) 58% 69.48% 66.667% 67.46%

Data Augmentation +Hog Features (1024,512,3) 68.51%. 77.63% 74.86% 75.69%
Data Augmentation +Hog Features (2048,1024,3) 68.525% 79.005 78.72% 78.45%
Data Augmentation + Hog Features + Hyper - 79.01% (GS) 71.2% -
parameter Tuning (1024,512,3) 78.825%(RF)
Data augmentation+ GLRLM features  63.54% 67.40% 65.46% 68.23%
Data Augmentation + Hog Features + Hyper - 80.04% 79.08% -
parameter Tuning(2048,1024,3)
 CONCLUSION
Ischemic stroke is the most common type of stroke and it is developed due to the blockage in the arteries
around the brain. Identifying the blood clot in the brain and predicting the type of stroke plays a vital role in
the diagnosis and treatment of various cerebrovascular diseases. The classification models address this
problem. These large ischemic stroke tiff images are initially compressed and reduced to a standard size
inorder to overcome the memory overload during the training of models.Form the results, it is observed
that the Random Forest Classifier with hyperparameter tuning gives good accuracy(80.4%) when compared
with the other models like Decision tree, XGBoost. In this work, the above mentioned models are applied to
the imbalanced data as well as the balanced data. The im-balanced data was converted into balanced with
the help of SMOTE augmentation technique. The performance of the model is increases in the balanced
dataset. The Transfer learning models like EfficientB7(67.4%) and VGG16(67.4%) gives moderate results. 
 REFERENCES
[1].Zhang, R., Zhao, L., Lou, W., Abrigo, J. M., Mok, V. C., Chu, W. C., ... & Shi, L. (2018). Automatic segmentation of acute ischemic stroke
from DWI using 3-D fully convolutional DenseNets. IEEE transactions on medical imaging, 37(9), 2149-2160. 
[2].C. -L. Chin et al., "An automated early ischemic stroke detection system using CNN deep learning algorithm," 2017 IEEE 8th International
Conference on Awareness Science and Technology (iCAST), 2017, pp. 368-372, doi: 10.1109/ICAwST.2017.8256481.
[3].K. C. Ho, W. Speier, H. Zhang, F. Scalzo, S. El-Saden and C. W. Arnold, "A Machine Learning Approach for Classifying Ischemic Stroke
Onset Time From Imaging," in IEEE Transactions on Medical Imaging, vol. 38, no. 7, pp. 1666-1676, July 2019, doi: 10.1109/TMI.2019.2901445.
[4] Kamal, H., Lopez, V., & Sheth, S. A. (2018). Machine learning in acute ischemic stroke neuroimaging. Frontiers in neurology, 9, 945.

[5] Kim, C., Zhu, V., Obeid, J., & Lenert, L. (2019). Natural language processing and machine learning algorithm to identify brain MRI reports
with acute ischemic stroke. PloS one, 14(2), e0212778.
[6] Nishio, M., Koyasu, S., Noguchi, S., Kiguchi, T., Nakatsu, K., Akasaka, T., ... & Itoh, K. (2020). Automatic detection of acute ischemic stroke
using non-contrast computed tomography and two-stage deep learning model. Computer Methods and Programs in Biomedicine, 196, 105711.
[7] Van Os, H. J., Ramos, L. A., Hilbert, A., Van Leeuwen, M., Van Walderveen, M. A., Kruyt, N. D., ... & Mr Clean Registry Investigators.
(2018). Predicting outcome of endovascular treatment for acute ischemic stroke: potential value of machine learning algorithms. Frontiers in
neurology, 9, 784..
[8]. Ho, K. C., Speier, W., Zhang, H., Scalzo, F., El-Saden, S., & Arnold, C. W. (2019). A machine learning approach for classifying ischemic
stroke onset time from imaging. IEEE transactions on medical imaging, 38(7), 1666-1676.
[9].M. Monteiro et al., "Using Machine Learning to Improve the Prediction of Functional Outcome in Ischemic Stroke Patients," in IEEE/ACM
Transactions on Computational Biology and Bioinformatics, vol. 15, no. 6, pp. 1953-1959, 1 Nov.-Dec. 2018, doi: 10.1109/TCBB.2018.2811471.

[10]. R. Su et al., "autoTICI: Automatic Brain Tissue Reperfusion Scoring on 2D DSA Images of Acute Ischemic Stroke Patients," in IEEE Transactions
on Medical Imaging, vol. 40, no. 9, pp. 2380-2391, Sept. 2021, doi: 10.1109/TMI.2021.3077113.

[11]. L. Tomasetti, L. J. Høllesli, K. Engan, K. D. Kurz, M. W. Kurz and M. Khanmohammadi, "Machine Learning Algorithms Versus Thresholding to
Segment Ischemic Regions in Patients With Acute Ischemic Stroke," in IEEE Journal of Biomedical and Health Informatics, vol. 26, no. 2, pp. 660-672,
Feb. 2022, doi: 10.1109/JBHI.2021.3097591.

[12]. V. L. Coli et al., "Detection of Brain Strokes Using Microwave Tomography," 2018 IEEE International Symposium on Antennas and Propagation
& USNC/URSI National Radio Science Meeting, 2018, pp. 223-224, doi: 10.1109/APUSNCURSINRSM.2018.8609404.

[13] Pathanjali C , Monisha G , Priya T , Ruchita Sudarshan K, Samyuktha Bhaskar, 2020, Machine Learning for Predicting Ischemic Stroke,
INTERNATIONAL JOURNAL OF ENGINEERING RESEARCH & TECHNOLOGY (IJERT) Volume 09, Issue 05 (May 2020).

[14].Li S, Zheng J, Li D. Precise segmentation of non-enhanced computed tomography in patients with ischemic stroke based on multi-scale U-Net deep
network model. Comput Methods Programs Biomed. 2021 Sep;208:106278. doi: 10.1016/j.cmpb.2021.106278. Epub 2021 Jul 9. PMID: 34274610.

[15].Yifan Hao, Aiyan Ji, Rongrong Xing, Wenqing Zhu, Baohong Jiang, Yi Jian, Hao Chen, Capillaries segmentation of NIR-II images and its
application in ischemic stroke, Computers in Biology and Medicine, Volume 147,2022,105742, ISSN 0010-4825.
[17] To, M.N.N., Kim, H.J., Roh, H.G. et al. Deep regression neural networks for collateral imaging from dynamic susceptibility contrast-enhanced magnetic resonance
perfusion in acute ischemic stroke. Int J CARS 15, 151–162 (2020). https://doi.org/10.1007/s11548-019-02060-7

[18] Amish Kumar, Neha Upadhyay, Palash Ghosal, Tamal Chowdhury, Dipayan Das, Amritendu Mukherjee, Debashis Nandi, CSNet: A new DeepNet framework for
ischemic stroke lesion segmentation, Computer Methods and Programs in Biomedicine, Volume 193,2020,105524,ISSN 0169-2607.

[19] L. Zhang et al., "Ischemic Stroke Lesion Segmentation Using Multi-Plane Information Fusion," in IEEE Access, vol. 8, pp. 45715-45725, 2020, doi:
10.1109/ACCESS.2020.2977415.

[20] Bin Zhao, Zhiyang Liu, Guohua Liu, Chen Cao, Song Jin, Hong Wu, Shuxue Ding, "Deep Learning-Based Acute Ischemic Stroke Lesion Segmentation Method on
Multimodal MR Images Using a Few Fully Labeled Subjects", Computational and Mathematical Methods in Medicine, vol. 2021, Article ID 3628179, 13 pages, 2021.
https://doi.org/10.1155/2021/3628179 

[21] Samak, Z. A., Clatworthy, P., & Mirmehdi, M. (2022). FeMA: Feature matching auto-encoder for predicting ischaemic stroke evolution and treatment
outcome. Computerized Medical Imaging and Graphics, 99, 102089.

[22] Zhao, B., Liu, Z., Liu, G., Cao, C., Jin, S., Wu, H., & Ding, S. (2021). Deep learning-based acute ischemic stroke lesion segmentation method on multimodal MR
images using a few fully labeled subjects. Computational and Mathematical Methods in Medicine, 2021.

[23] Platscher, M., Zopes, J., & Federau, C. (2022). Image translation for medical image generation: Ischemic stroke lesion segmentation. Biomedical Signal Processing
and Control, 72, 103283.
[24] Islam, M., Vaidyanathan, N. R., Jose, V., & Ren, H. (2018, September). Ischemic stroke lesion segmentation using adversarial learning. In International
MICCAI Brainlesion Workshop (pp. 292-300). Springer, Cham.

[25] Kappelhof, N., Ramos, L. A., Kappelhof, M., van Os, H. J., Chalos, V., van Kranendonk, K. R., ... & Marquering, H. A. (2021). Evolutionary algorithms and
decision trees for predicting poor outcome after endovascular treatment for acute ischemic stroke. Computers in Biology and Medicine, 133, 104414.

[26] Zhang, S., Wang, J., Pei, L., Liu, K., Gao, Y., Fang, H., ... & Xu, Y. (2022). Interpretable CNN for ischemic stroke subtype classification with active model
adaptation. BMC medical informatics and decision making, 22(1), 1-12.

[27] Shi, T., Jiang, H., & Zheng, B. (2021). C 2 MA-Net: Cross-Modal Cross-Attention Network for Acute Ischemic Stroke Lesion Segmentation Based on CT Perfusion
Scans. IEEE Transactions on Biomedical Engineering, 69(1), 108-118.

[28]. H. Kuang, B. K. Menon and W. Qiu, "Segmenting Hemorrhagic and Ischemic Infarct Simultaneously From Follow-Up Non-Contrast CT Images in Patients With
Acute Ischemic Stroke," in IEEE Access, vol. 7, pp. 39842-39851, 2019, doi: 10.1109/ACCESS.2019.2906605. 

[29] Clèrigues, A., Valverde, S., Bernal, J., Freixenet, J., Oliver, A., & Lladó, X. (2019). Acute ischemic stroke lesion core segmentation in CT perfusion images using
fully convolutional neural networks. Computers in biology and medicine, 115, 103487

[30] Zhao, B., Ding, S., Wu, H., Liu, G., Cao, C.,Jin, S., & Liu, Z. (2019). Automatic acute ischemic stroke lesion segmentation using semi-supervised
learning. arXiv preprint arXiv:1908.03735. 
THANK YOU