The Advancements of Convolutional Neural Networks on Cerebral

Hemorrhage Detection

Xiao Liu

Department of Software Engineering, Tongji University, Shanghai, China

Keywords: Cerebral Hemorrhage, Deep Learning, Convolutional Neural Networks.

Abstract: Cerebral hemorrhage is a common and serious disorder that poses a serious threat to the health of the patient.

Due to the shortcomings, such as the low efficiency of traditional cerebral hemorrhage detection, it is rather

necessary to consider techniques combined with artificial intelligence to enhance the quality and speed of

detection because of the intractability of the disease. In this paper, a method using convolutional neural

networks (CNN) is considered, studied, and further discussed. Currently, deep-learning-based automated

cerebral hemorrhage detection methods have gained widespread attention. These approaches have achieved

rapid and accurate brain bleeding detection by analyzing head imaging data, such as computerized

tomography (CT) images. Some professors adopted a special technique or structure, for example, the attention

mechanism or hybrid CNN, to detect and classify the CT images, which has already gained wonderful

achievements. The use of attention mechanisms or mixed CNN for brain hemorrhage testing contributes to

improving the accuracy, adaptability, and efficiency of testing, which is one of the important directions of

current research. However, in practical applications, some models have been poorly performed in dealing

with specific types of brain bleed and have limited generalization capabilities. The focus in this field includes

improving character representation, optimizing model structures, and solving data deviations to improve the

generalizing capability and accuracy of models. In conclusion, this paper provides a good overview of cerebral

hemorrhage detection.

1 INTRODUCTION

Cerebral hemorrhage refers to when cerebrovascular

rupture or vascular wall problems occur and blood

flows into brain tissue, causing increased stress and

causing brain tissues to be damaged or even killed.

The symptoms of cerebral bleeding may include

severe headaches, nausea, vomiting, awareness loss,

drainage, and body impotence (Peng, 2019; Kumar,

2023). Cerebral hemorrhage was one kind of disease

that seriously jeopardized the health of human beings.

Traditional brain bleeding is usually diagnosed

using the following methods: the clinic will first

consider the possibility of cerebral bleeding with the

patient's past medical history and clinical symptoms,

and then with a Computerized Tomography (CT) of

the scrotum to make a clear diagnosis (Al'Aref, 2019;

Bloom, 1996). It can be observed that the traditional

detection method has certain drawbacks. For

example, due to the need for a doctor's judgment and

https://orcid.org/0009-0006-1857-264X

CT detection, the detection time is long, and the rate

is low, and in addition, misdiagnosis is easy to occur.

Besides, the labor cost of the whole detection

procedure is high. Therefore, it is necessary to

combine the traditional detection technique with

Artificial Intelligence (AI). Initially, AI can process

large amounts of medical image data and perform

analysis and diagnosis in a short time (Qiu, 2022). In

addition, through deep learning and pattern

recognition technology, AI can be trained to learn

from a large amount of image data, and then AI can

extract key features from it, thus making accurate

prediction (Sun, 2020; Zhou, 2023). Therefore, the

accuracy of AI can help reduce the risk of

misdiagnosis and missed diagnosis and improve the

diagnostic accuracy of cerebral hemorrhage (Wang,

2013).

In recent years, there have been many significant

advances in the field of artificial intelligence,

especially for deep learning algorithms. Deep

306

Liu, X.

The Advancements of Convolutional Neural Networks on Cerebral Hemorrhage Detection.

DOI: 10.5220/0012937200004508

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 306-310

ISBN: 978-989-758-713-9

learning (DL), for example, which uses neural

networks to simulate the structure and function of the

human brain for learning and pattern recognition of

large-scale data, has achieved great success in image

recognition, speech recognition, natural language

processing, and other fields such as Convolutional

Neural Networks (CNN) and Recurrent Neural

Networks (RNN). With the rapid development of AI,

it has been used in various fields such as engineering

(Pham, 1999), education (Beck, 1996), agriculture

(Eli-Chukwu, 2019) etc. It is worth noting that AI

technology is also playing an important role in

medical fields such as gastroenterology (Yang, 2019),

radiology (Teramoto, 2019), and cardiology

(Johnson, 2018). AI is usually used in medical image

processing (Wang, 2013; Castiglioni, 2021), data

storage, targeted therapy (Haleem, 2019) etc. For

instance, AI can help detect and diagnose pneumonia

because CNNs have the ability to classify medical

images (Li, 2023; Račić, 2021). Although there are

relatively few studies on the combination of cerebral

hemorrhage and AI compared to other fields, there are

still some achievements proposed. Lee et al.

developed a novel deep-learning algorithm for

artificial neural networks (ANN) to evaluate its

feasibility for detecting Intracranial Hemorrhage

(ICH) and classifying its subtypes (Lee, 2020). Zhang

et al. segmented the brain parenchyma area using the

Mask Region-based Convolutional Neural Network

(Mask R-CNN network). Then they located the blood

clot area using the threshold segmentation approach.

Finally, cross-sectional contour interpolation was

used to construct a 3D visualization technique for

cerebral bleeding (Zhang, 2021). Due to the

importance of this area, deep learning, especially

CNN, has made significant breakthroughs in brain

hemorrhage in recent years, it is therefore necessary

to make a comprehensive overview of it.

2 METHOD

In the biomedical field, machine learning has had a

significant impact on the prediction and detection of

cerebral hemorrhage. Machine learning can help

facilitate the identification and prediction of diseases

of concern in the medical industry, and perhaps even

the fairness of decision-making (Bharath Kumar

Chowdary, 2022).

2.1 Framework of CNN-Based

Hemorrhage Detection

Figure 1 presents the concrete process of CNN-based

hemorrhage detection, which typically include the

following steps: data collection, data preprocessing,

model building, model training, testing and

assessment, and deployment.

Figure 1: The structure (Picture credit: Original).

Data Collection: A large number of medical

imaging data sets, such as CT scans or Magnetic

Resonance Imaging (MRI) images, should be

collected from hospitals or medical institutes.

Data Pre-Processing: This may include

adjusting the size of the image, usually scaling it to a

uniform size. Data pre-processing improves data

quality and reduces noise interference through

operations such as cleaning, converting, and

naturalizing raw data, thereby helping machine

learning models to learn and generalize better. For

example, localization improves the stability of model

training, increases the convergence rate, and

effectively solves the problem of gradient

disappearance or explosion. Data enhancement can

scale up data sets, improve the generalization

capacity of models, etc. Through these data

preprocessing techniques, models can be trained,

generalized, and robust, thereby better addressing

actual problems and improving the performance of

machine learning models.

Model Construction: CNN is a deep neural

network consisting of the convolutional layer, the

pooling layer, and the fully connected layer. The

convolutional layer is used to extract the

characteristics of an image and generate a series of

characteristics by rolling the convolution kernel onto

the image. The pooling layer is used to reduce the

dimensions of a feature map, reduce the number of

calculations, and retain the important characteristics.

The Fully Connected Layer maps the features

obtained by the Pooling Layer and connects the

Output Layer for the final classification or regression

task.

The Advancements of Convolutional Neural Networks on Cerebral Hemorrhage Detection

307

Training and Assessment: The next step is to

train the built-in CNN model using the ready-made

data set. It is also important to use a separate test set

to evaluate trained CNN models, with indicators such

as accuracy, recall rate, precision, and F1 scores

assessing the performance of models.

Deployment: once the model has been evaluated

and reached the required performance indicators, it

can be deployed into practical applications.

2.2 CNN Combined with Attention

Mechanisms

Attention mechanisms are used in neural networks to

focus on specific parts of the input sequence or image,

enabling models to attach greater importance to

relevant information.

Alis, D. et al. employed a unique DL architecture,

a hybrid CNN recurrent neural network (RNN) with

an attention mechanism, to detect and subcategorize

ICH on non-contrast head CT images (Alis, 2022).

Initially, continuous, uncontrolled enhanced CT

scans of the head are obtained from the emergency

departments of five tri-methyl clinics. Then Five

neuroscientists evaluate the images collected to

determine the presence of bleeding and, if any, mark

their subtypes (intraparenchymal hemorrhage (IPH),

intraventricular hemorrhage (IVH), subdural

hematoma (SDH), epidural hematoma (EDH), and

subarachnoid hemorrhage (SAH)). Using the

TensorFlow deep learning library, a joint CNN-RNN

model is constructed on a customized workstation.

The model uses InceptionResNetV2 as the basic

network to extract the most relevant features of the

image. The extracted images are delivered to a two-

way RNN with a focus layer to deliver information

between images. The attention mechanism helps to

concentrate the most pertinent data needed by the

two-wheel RNN to focus on the task. It uses three

different window-point settings to emphasize contrast

differences between background and ICH and

performs some conventional image pre-processing

operations before delivering the image to the

network. After that, these professors divide the data

set into training sets and validation sets, with four

centers of data used for the training model and one

center for the validation model. Besides, an improved

NormGrad method has been used to generate a

Gradient-weighted Class Activation Mapping (Grad-

CAM) to highlight the decision-based basis of the

model on a given task.

A. Hussain et al. propose a novel deep learning-

based CNN model to efficiently detect and classify

brain hemorrhage and its subtypes. This paper

describes a hybrid attention-based ResNet

architecture for ICH detection and classification

(Hussain, 2022).

The study used 434,166 CT scan samples from the

RNSA-2019 dataset, including five separate ICH

categories. Then researchers preprocess these CT

images. They used the Deep Volume Generation

Confrontation Network (DCGAN) to generate CT

scan images of the epidural hemorrhage (EH)

category to address data set category imbalances. In

addition, feature extraction is done using the

attention-based ResNet-152V2 architecture. Feature

selection, redundancy removal, and de-

dimensionation operations are performed using

master component analysis. Besides, they used the

gradient enhancement algorithm XGBoost. The next

step is to optimize the model's learning process by

performing super parametric adjustments manually to

improve its performance. After optimizing the model,

experiments with Python programming languages in

Kaggle Jupyter Notebook were carried out to create

deep learning models and evaluate their performance.

Finally, some indicators, including accuracy,

precision, recall rate, F1 score, true rate, true

negative, AUC, etc., were used to evaluate the

performance of the model.

2.3 Hybrid CNN

Hybrid CNN hemorrhage detection methods enhance

the accuracy, robustness, and generalization of

intracranial bleeding detection by combining

different types of CNN models.

For example, Iqbal et al. use hybrid machine

learning algorithm to detect brain hemorrhage.

The study used a CT brain imaging data set from

Kaggle as input. Then the study selected different 3D

volume neural network (3D VNN) models, including

Visual Geometry Group-16 (VGG-16), Visual

Geometry Group-19 (VGG-19), etc., as well as other

models such as the Multilayer Perceptron Model

(MLP), Support Vector Machine (SVM), and

Random Forest (RF). Besides, hybrid machine

learning algorithms are designed by combining

different 3D VNN models (such as VGG-16 and

VGG-19) with Random Forest and Multilayer

Perceptron (MLP) classifiers. It uses the selected

model to train CT brain images and test the accuracy

of the model. Based on the results, the authors

claimed that the combined method of the VGG-16

and MLP classifiers achieved an optimal accuracy of

about 97.24%. The study also uses explanatory AI

technology to explain the model's predictions for the

brain hemorrhage category, making the predictive

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

308

process of the model more understandable (Iqbal,

2022).

3 DISCUSSION

By analyzing current research, CNN can indeed

improve the efficiency of the detection process.

However, there are some limitations and challenges

in current studies that influence the promotion of this

method.

Initially, for example, there may be imbalances in

brain and non-brain bleeding samples, resulting in

poor performance of models in a few categories.

Moreover, medical imaging data may be influenced

by realistic factors, so the data needs to be cleaned up

and preprocessed.

Secondly, it is necessary to consider the

interpretability and domain knowledge of models to

better serve doctors and patterns. Despite the use of

interpretable AI technologies to interpret predictions,

the interpretability of models remains a challenge for

the medical field (Iqbal, 2022). The complexity of

some deep learning models may limit their

interpretability, which requires weighing the

relationship between model performance and

interpretability. Besides, as for the domain knowledge,

the need for brain hemorrhage testing in different

clinical scenarios may vary, so targeted models and

solutions are needed, which is also a deficiency of

current research.

Thirdly, model generalization is also a very

important factor. Model-trained datasets may differ

from actual clinical data and require consideration of

how to enhance the model's generalization

capabilities across different data sets. Inadequate

generalization may be due to the limitations of the

training data set, such as the lack of diversity of data

and the inability to cover various types of cerebral

bleeding. In addition, in terms of characteristic

representation, the various manifestations of brain

bleeding may not be captured. These factors can

result in the poor performance of the model for

different types of cerebral hemorrhage.

But in the future, through some methods and

research, these issues may be improved continuously.

Here are some possible solutions.

Firstly, it is possible to combine Shapley Additive

Explanations (SHAP), a method for explaining model

predictions that can help to understand the degree of

the model's contribution to the input characteristics,

values, and visualization techniques, such as

generating thermal charts or local significance charts,

to visually present the model's critical characteristics

and diagnostic basis for brain hemorrhage detection,

enhancing the interpretability and credibility of the

model.

Secondly, transfer learning can be used. Using

transfer learning methods such as feature extraction

and model fine-tuning, existing brain hemorrhage

detection models are applied to new datasets and

adjusted accordingly to the characteristics of the new

data sets to improve the performance of the model in

the new dataset (Qiu, 2019).

Thirdly, the use of field-specific techniques, such

as field counter-training and field-to-field distance

measurement, enables the model to fully adapt to the

data characteristics of different medical institutions

and improves its generalization capacity, which is

also a possible solution.

Furthermore, different options might be examined

from other perspectives on expert systems and federal

learning. Developing an expert system with medical

specialists to codify expert information using

knowledge graphs, rules engines, and other

technologies, as well as to deliver more complete,

accurate brain hemorrhage testing and diagnosis

support using machine learning models, Creating a

federal learning framework, utilizing technologies

such as secure multifaceted computing, differential

privacy, sharing and updating model parameters

across several medical institutions, and enhancing the

performance and applicability of brain hemorrhage

testing models.

4 CONCLUSIONS

As the research has demonstrated, investigations have

indicated that CNN has produced exceptional results

in cerebral hemorrhage diagnosis, displaying

excellent accuracy and sensitivity in the field of

medical imaging analysis. CNN technology can assist

clinicians in enhancing the efficiency and correctness

of their diagnoses in automated cerebral hemorrhage

detection. As a result, it can play an important role in

clinical sectors such as supporting doctors in rapidly

and properly diagnosing illnesses, initiating early

treatment, and lowering patient death and disability

rates. However, there are still constraints that need to

be investigated further, such as the poor quality of the

data, the model interpretability, domain knowledge,

and model generalization. It only covers the CNN

algorithm for detecting brain hemorrhages, which has

certain drawbacks. Indeed, other types of algorithms

should be investigated in the future. As a result, CNN

technology must be constantly enhanced in order to be

The Advancements of Convolutional Neural Networks on Cerebral Hemorrhage Detection

309

suitable for wider clinical use and diffusion in the

future.

REFERENCES

Alis, D., Alis, C., Yergin, M., Topel, C., Asmakutlu, O.,

Bagcilar, O., ... & Karaarslan, E. 2022. A joint

convolutional-recurrent neural network with an

attention mechanism for detecting intracranial

hemorrhage on noncontrast head CT. Scientific

Reports, 12(1), 2084.

Al'Aref, S. J., & Min, J. K. 2019. Cardiac CT: current

practice and emerging applications. Heart (British

Cardiac Society), 105(20), 1597–1605.

Beck, J., Stern, M., & Haugsjaa, E. 1996. Applications of

AI in Education. XRDS: Crossroads, The ACM

Magazine for Students, 3(1), 11-15.

Bharath Kumar Chowdary, P., Jahnavi, P., Rani, S. S.,

Chowdary, T. J., & Srija, K. 2022. Detection and

Classification of Cerebral Hemorrhage Using Neural

Networks. In Proceedings of Second International

Conference on Advances in Computer Engineering and

Communication Systems: ICACECS 2021 (pp. 555-

564). Singapore: Springer Nature Singapore.

Bloom, A. I., Neeman, Z., Floman, Y., Gomori, J., & Bar-

Ziv, J. 1996. Occipital condyle fracture and ligament

injury: imaging by CT. Pediatric radiology, 26, 786-

790.

Castiglioni, I., Rundo, L., Codari, M., Di Leo, G., Salvatore,

C., Interlenghi, M., ... & Sardanelli, F. 2021. AI

applications to medical images: From machine learning

to deep learning. Physica medica, 83, 9-24.

Eli-Chukwu, N. C. 2019. Applications of artificial

intelligence in agriculture: A review. Engineering,

Technology & Applied Science Research, 9(4).

Haleem, A., Javaid, M., & Khan, I. H. 2019. Current status

and applications of Artificial Intelligence (AI) in

medical field: An overview. Current Medicine

Research and Practice, 9(6), 231-237.

Hussain, A., Yaseen, M. U., Imran, M., Waqar, M.,

Akhunzada, A., Al-Ja’afreh, M., & El Saddik, A. 2022.

An attention-based ResNet architecture for acute

hemorrhage detection and classification: Toward a

Health 4.0 digital twin study. IEEE Access, 10, 126712-

126727.

Iqbal, K. N., Azad, I., Emon, M. I. H., Amlan, N. S., &

Aporna, A. A. 2022. Brain hemorrhage detection using

hybrid machine learning algorithm (Doctoral

dissertation, Brac University).

Johnson, K. W., Torres Soto, J., Glicksberg, B. S., Shameer,

K., Miotto, R., Ali, M., ... & Dudley, J. T. 2018.

Artificial intelligence in cardiology. Journal of the

American College of Cardiology, 71(23), 2668-2679.

Kumar, E. A., Nikitha, P., & Bai, P. J. 2023. Clinical profile

of spontaneous cerebellar hemorrhage-An original

article. International Archives of Integrated Medicine

10(4).

Lee, J. Y., Kim, J. S., Kim, T. Y., & Kim, Y. S. 2020.

Detection and classification of intracranial

haemorrhage on CT images using a novel deep-learning

algorithm. Scientific reports, 10(1), 20546.

Li, X., Sun, Z., & Zhang, L. 2023. Research advances of

artificial intelligence-based medical imaging in the

screening, diagnosis and prediction of pneumonia.

Journal of Shandong University (Health Sciences),

61(12), 13-20. (in Chinese)

Peng, Q., Li, H. L., Wang, Y., & Lu, W. L. 2019. Changing

trend regarding the burden on cerebrovascular diseases

between 1990 and 2016 in China. Zhonghua liu Xing

Bing xue za zhi= Zhonghua Liuxingbingxue Zazhi,

40(4), 400-405.

Pham, D. T., & Pham, P. T. N. 1999. Artificial intelligence

in engineering. International Journal of Machine Tools

and Manufacture, 39(6), 937-949.

Qiu, Y., Chang, C. S., Yan, J. L., Ko, L., & Chang, T. S.

2019. Semantic segmentation of intracranial

hemorrhages in head CT scans. In 2019 IEEE 10th

International Conference on Software Engineering and

Service Science (ICSESS) (pp. 112-115). IEEE.

Qiu, Y., Wang, J., Jin, Z., Chen, H., Zhang, M., & Guo, L.

2022. Pose-guided matching based on deep learning for

assessing quality of action on rehabilitation

training. Biomedical Signal Processing and Control, 72,

103323.

Račić, L., Popović, T., & Šandi, S. 2021. Pneumonia

detection using deep learning based on convolutional

neural network. In 2021 25th International Conference

on Information Technology (IT) (pp. 1-4). IEEE.

Sun, G., Zhan, T., Owusu, B.G., Daniel, A.M., Liu, G., &

Jiang, W. 2020. Revised reinforcement learning based

on anchor graph hashing for autonomous cell activation

in cloud-RANs. Future Generation Computer Systems,

104, 60-73.

Teramoto, A. 2019. Application of artificial intelligence in

radiology. Gan to Kagaku ryoho. Cancer &

Chemotherapy, 46(3), 418-422.

Wang, Y., & Li, C. 2013. Recent advances in the

application of artificial intelligence in medical image

processing. Chinese Journal of Medical Physics,

30(003), 4138-4143. (in Chinese)

Yang, Y. J., & Bang, C. S. 2019. Application of artificial

intelligence in gastroenterology. World journal of

gastroenterology, 25(14), 1666.

Zhang, T., Song, Z., Yang, J., Zhang, X., & Wei, J. 2021.

Cerebral hemorrhage recognition based on Mask R-

CNN network. Sensing and Imaging

, 22(1), 1.

Zhou, Y., Osman, A., Willms, M., Kunz, A., Philipp, S.,

Blatt, J., & Eul, S. 2023. Semantic Wireframe

Detection. publica.fraunhofer.de.

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

310