Automatic Approach for Brain Aneurysm Detection Using Convolutional Neural Networks

Abstract

The paper introduces an approach for detecting brain aneurysms, a critical medical condition, by utilizing a combination of 3D convolutional neural networks (3DCNNs) and Convolutional Long Short-Term Memory (ConvLSTM). Brain aneurysms pose a significant health risk, and early detection is vital for effective treatment. Traditional methods for aneurysm detection often rely on complex and time-consuming procedures. A radiologist specialist annotates each aneurysm and supports our work with true-ground annotations. From the annotated data, we extract images to train proposed neural networks. The paper experiments with two different types of networks, specifically focusing on 2D convolutional neural networks (2DCNNs), 3D convolutional neural networks (3DCNNs), and Convolutional Long Short-Term Memory (ConvLSTM). Our goal is to create a virtual assistant to improve the search for aneurysm locations, with the aim of further realizing the virtual assistant. Subsequently, a radiologist specialist will confirm or reject the presence of an aneurysm, leading to a reduction in the time spent on the searching process and revealing hidden aneurysms. Our experimental results demonstrate the superior performance of the proposed approach compared to existing methods, showcasing its potential as a valuable tool in clinical settings for early and accurate brain aneurysm detection. This innovative fusion of 3DCNN and LSTM (3DCNN-ConvLSTM) techniques not only improves diagnostic precision but also holds promise for advancing the field of medical image analysis, particularly in the domain of neurovascular diseases. Overall, our research underscores the potential of neural networks for the machine detection of brain aneurysms.

1. Introduction

In recent years, deep learning approaches, particularly convolutional neural networks (CNNs), have demonstrated significant potential for detecting brain aneurysms from medical images. In our research, we aim to explore the use of 2DCNNs for aneurysm detection, with the goal of enhancing the accuracy and efficiency of existing detection methods. A brain aneurysm refers to a bulge or ballooning in a blood vessel in the brain, often resembling a berry hanging on a stem. If an aneurysm ruptures, it can lead to bleeding in the brain, becoming a life-threatening condition that requires prompt medical treatment. However, most brain aneurysms do not rupture, cause health problems, or exhibit symptoms. These aneurysms are typically detected incidentally during tests conducted for other medical conditions. In certain instances, treatment for an unruptured brain aneurysm may be deemed appropriate and could potentially prevent a rupture in the future. Brain aneurysms are enormous threats to human health, with a prevalence of approximately 4% [1,2]. The rupture of an aneurysm usually causes death or severe damage to the patients.

Brain aneurysms are potentially life-threatening conditions that require early detection and proper treatment to prevent serious consequences such as stroke, hemorrhage, or even death [3]. Brain aneurysm detection from medical images such as computed tomography (CT) and magnetic resonance imaging (MRI) scans is challenging. It requires accurate and reliable analysis of complex data. Both MRI and CT scans can view internal body structures [4]. Nevertheless, a CT scan is quicker and can generate images of tissues, organs, and skeletal structures. On the other hand, an MRI excels at capturing detailed images that assist doctors in identifying the presence of abnormal tissues within the body. MRIs provide a higher level of detail in their scan images.

Deep neural networks (DNNs) have emerged as promising approaches for detecting brain aneurysms. DNNs are machine learning algorithms capable of automatically learning and extracting relevant features from large datasets, enabling accurate and efficient analysis of medical images. In this article, we will delve into the state-of-the-art approaches in brain aneurysm detection using DNNs, including convolutional neural networks (CNNs), transfer learning, and ensembling techniques. We will also discuss the various types of medical images employed for brain aneurysm detection and the potential benefits of using DNNs for segmentation. Overall, this article aims to provide an overview of the latest developments in brain aneurysm detection using DNNs and their potential to enhance patient outcomes.

The detection of aneurysms is a time-consuming and technically demanding task. It uses a computed tomography aneurysm (CTA) machine. It uses X-rays to scan parts of the human body from multiple angles and multiple spatial projections, respectively. Yoon compares different CTA-based aneurysm analysis techniques [5]. It is laborious and time-consuming work. The radiologist specialist must analyze hundreds of images carefully for every single patient. He must explore details from different views to determine whether they are normal blood vessels or contain aneurysms. However, the small size of aneurysms and low-intensity contrast to normal vessels requires a sub-specially trained radiologist specialist.

Bo et al. [6] investigated how computer-assisted detection of intracranial aneurysms (IAs) using a deep neural network can improve the efficiency and speed of analysis. He utilized the GLIA-Net model for CTA image segmentation. His research showed that Artificial Inference (AI) detects aneurysms approximately 6 times faster than specialized radiologists. The low accuracy of the automatic detection—it hovered around 20%, while a human worked with an accuracy of around 90%—emerged as a shortcoming of AI. AI was able to achieve an average recall of 85% on average. The two radiologists reached slightly better sensitivity of 86% and 88%, while the average value was around 70% for the human. In each case, the AI-assisted analysis became faster. Soun et al. [7] focused on using AI for acute stroke imaging. In [8], the detection of a pulmonary nodule using multilevel 3D CNN is presented.

Aneurysms can also be detected using magnetic resonance angiography (MRA). The MRA-based analysis is more time-consuming than CTA but can create a precise three-dimensional vascular system model. Park et al. [9] used deep learning-assisted diagnosis of cerebral aneurysms using the HeadXNetModel. The latter works on the principle of a three-dimensional convolutional neural network. He focused on measuring sensitivity, specificity, accuracy, and time. He compared the radiologist’s work performance with and without assisted analysis. It showed that assisted analysis improved the performance of all participants (six radiologists, a neurosurgeon, and a resident). Radiologist 5 performed the analysis most accurately of all participants. The assisted search did little to improve his performance. It slightly worsened his specificity. Radiologist 3 had the poorest results with unaugmented analysis, but with the augmented analysis, he achieved comparable results to more experienced colleagues.

Also, the COVID-19 pandemic has brought with it new challenges for telemedicine. Bechstein [10] describes his first experience with intracranial treatment using a dedicated live-streaming technology. Meng et al. [11] experimented with a multi-modal convolutional neural network. He filtered the original CTA images using Gaussian and Laplace and co-presented the original image with multiple modalities at the model input. Dai et al. [12] analyzed data from 208 patients who had 222 diagnosed aneurysms using DNN. They had them categorized according to size and location. They also evaluated the cases of false positives and false negatives. In [13,14], they compared the work of several teams using different methods and approaches to detect aneurysms in the context of the Adam challenge on 3D data. Stember et al. [15] investigated CNNs on MRI images for unusual aneurysms of different sizes.

Zhou et al. [16] also used a 3D-CNN-based spatial model. They compared the accuracy of different types of models based on DNN principles. Zeng et al. [17] used an automatic diagnosis approach based on the spatial information of fusion features. Hou et al. [18] also performed a similar study, but they proposed a 1D-CNN and compared their results with other authors.

Nowadays, artificial intelligence (AI) is increasingly used in the management of patients with cerebrovascular disease, especially in the case of Cerebral Misery Perfusion (CMP stroke). AI assists physicians and healthcare professionals in identifying risk factors and predicting treatment outcomes. In addition, AI also helps in diagnosis, for example, in identifying possible symptoms of stroke and determining their severity. Another improvement that AI brings in the management of CMP patients is aneurysm detection [19]. An aneurysm is a condition where there is a bulge in the wall of a blood vessel that can cause bleeding into the brain. Detection of an aneurysm is very important because early diagnosis can prevent serious consequences and save the patient’s life. The use of AI in aneurysm detection allows doctors to identify the symptoms of the condition by analyzing imaging scans such as CT scans or MRIs. The results of the analysis are then processed and interpreted using machine learning algorithms, allowing doctors to provide more accurate and faster diagnoses. The benefit of using AI in the management of CMP patients is primarily to improve diagnosis and prediction of treatment outcomes, leading to better patient health outcomes. It also allows doctors and healthcare professionals to work faster and more efficiently, leading to an improved overall healthcare system [20].

Our work and methods emerged from the data we had available. We focused on processing CTA images captured from a series of angles around the head. We tested different types of deep neural networks to optimize the accuracy of brain aneurysm detection. In the methodology, we describe in detail the form of the data used, aneurysm annotation, and pre-processing. We tried a two-dimensional convolutional neural network (2DCNN) model and a model based on a combination of 3DCNN and ConvLSTM (3DCNN-ConvLSTM).

In this study, we introduce a cutting-edge methodology that leverages 3DCNN to extract intricate 3D features from medical imaging data, providing a comprehensive representation of vascular structures. Furthermore, we incorporate ConvLSTM, a specialized recurrent neural network, to capture temporal dependencies in the data, enhancing the accuracy of aneurysm detection by considering dynamic changes over time. The novelty of combining 3D convolutional neural networks (3DCNNs) with Convolutional Long Short-Term Memory networks (ConvLSTM) lies in the enhanced capability to analyze spatiotemporal features within volumetric data. This fusion of architectures represents a significant advancement in the field of deep learning, particularly for applications dealing with three-dimensional sequences or video-like data. In summary, the novelty of combining 3D CNN with ConvLSTM lies in the holistic analysis of volumetric data, considering both spatial and temporal dimensions. This hybrid architecture is particularly powerful in tasks requiring a comprehensive understanding of dynamic three-dimensional sequences, making it a valuable tool in various domains, from video analysis to medical imaging.

2. Materials and Methods

Our research aims to investigate the usage of deep neural networks (DNNs) for aneurysm detection and compare their performance to existing detection methods. We will also explore different network architectures of DNN and training strategies to optimize the accuracy of the models. Overall, our goal is to develop a more accurate and efficient method for detecting brain aneurysms, which can ultimately improve patient outcomes. We focused on a 2D convolutional neural network model and a combination of 3DCNN and LSTM. Further, in this section, we will describe the entire process, including the dataset, annotation, signal pre-processing, and model design.

2.1. 2DCNN Architecture

The 2D convolutional neural networks (2DCNNs) are a type of deep learning architecture (see Figure 1) commonly used for image classification and detection tasks, including brain aneurysm detection. The network has multiple layers, including convolutional, pooling, and fully connected layers [15].

In the context of brain aneurysm detection, the 2DCNN takes in a 2D image of the brain, such as a computed tomography angiography (CTA) or magnetic resonance imaging (MRI) scan. The network then applies a series of convolutional filters to the image to extract features at different levels of abstraction. These features may include edges, corners, and other patterns that may indicate the presence of an aneurysm. After the convolutional layers, the outputs pass through pooling layers. They reduce the dimensionality of the feature maps while retaining the most salient features. This step helps improve the computational efficiency of the model while still preserving the essential information. The pooling layers’ output passes through a series of fully connected layers, which use the extracted features to predict the presence or absence of an aneurysm. The final layer typically uses a Sigmoid or Softmax activation function to produce a probability value for each class [15].

The 2DCNN performance for aneurysm detection is optimized by various modifications to the architecture. For example, skipping connections between the convolutional layers can better capture the spatial information in medical images. Using residual blocks, which allow for better gradient flow through the network, can also improve the model’s accuracy.

2.2. 3DCNN Architecture

The 3DCNN neural network is capable of analyzing and recognizing various dynamic 2D objects present in both images and 3D image datasets, such as those encountered in medical imaging applications. In the 3DCNN architecture, the 3D convolution operation (see Figure 2) is applied to the dataset along three directions (x, y, z) using a three-dimensional filter. It is important to note that the values within the layer of the three-dimensional filter must be configured to be non-negative [16]. The 3D convolution operates by arranging neighboring layers around the central cube, establishing interconnected convolutional maps to capture motion information. However, it is important to note that each convolutional kernel can only capture a specific type of feature.

Enhancing the performance of a 3DCNN, much like its 2D counterpart, involves combining multiple convolutional layers. When constructing a 3DCNN model, it becomes crucial to carefully configure parameters such as the number of layers, the quantity of filters in each layer, and the filter size. In cases where pooling is integrated into the neural network architecture, the pooling size must be defined in three dimensions to accommodate the 3D data. Consequently, the output shape produced by the 3DCNN network is represented as a three-dimensional volume space [17,18].

2.3. ConvLSTM Architecture

Convolutional Long Short-Term Memory (ConvLSTM) is a neural network type that merges the capabilities of both convolutional neural networks (CNNs) and Long Short-Term Memory (LSTM) models. The CNNs are conventionally applied to image processing, excelling at feature identification and extraction from images. Conversely, LSTM models find their utility in time-series analysis, employing memory cells to preserve information across time steps. ConvLSTM is specifically engineered to unify these two models, making it proficient at managing spatiotemporal data encompassing both spatial and temporal dependencies. The ConvLSTM model’s structure (see Figure 3) comprises four primary components [21]:

• Input Gate: The input gate plays a pivotal role in determining which new information should be incorporated into the memory cell. It takes input from the preceding time step and the current pixel value, generating an output ranging from 0 to 1. A value of 0 signifies that no new information should be incorporated, whereas a value of 1 signifies that all incoming information should be preserved.
• Forget Gate: The forget gate assumes the crucial role of determining which information should be discarded from the memory cell within the LSTM model. It receives input from both the previous time step and the current pixel value, generating an output within the range of 0 to 1. A value of 0 signifies that the model should erase the prior information, whereas a value of 1 denotes that the information should be preserved. Importantly, the forget gate factors in input data and the hidden state from the preceding time step in its decision-making process.
• Output Gate: The output gate assumes the role of producing the hidden state at the present time step. It considers input from both the previous time step and the current pixel value, along with the memory cell state, yielding an output within the range of 0 to 1. This output is subsequently multiplied by the hyperbolic tangent of the memory cell, resulting in the updated hidden state. Essentially, the output gate plays a crucial part in regulating the quantity of information to transmit to the subsequent time step.

Figure 3 illustrates the architecture of a ConvLSTM network. The $\sigma$ symbol represents the sigmoid function, tanh represents the hyperbolic tangent function, and t represents time step. The LSTM is equipped with three gates designed to capture long-term dependencies. These gates are the input gate, the output gate, and the forget gate. Each gate plays a crucial role in determining whether the LSTM should store or discard newly acquired information in its memory cell [21].

The ConvLSTM model represents a robust neural network framework that seamlessly integrates convolutional and LSTM models, allowing it to effectively manage spatiotemporal data. By preserving both spatial and temporal information, this model demonstrates remarkable prediction accuracy in various domains, including bioinformatics, traffic prediction, and video processing. Its versatility and effectiveness have been showcased through its successful application in numerous real-world scenarios [21].

3. Description of the Dataset

We work with a small database of anonymized Digital Subtraction Angiography (DSA) from our research partner at the Department of Radiology at the University Hospital in Martin, Slovak Republic (33 patients), as shown in Figure 4. It contains angiograms from patients affected by different cerebral aneurysms. Each patient’s record is in Digital Imaging and Communications in Medicine (DICOM) format. Every DICOM record has 122 X-ray images of the head from rotating angles. Each image has a resolution of 1024 × 1024 pixels. The DICOM dataset pixels use 16-bit unsigned integer coding—grayscale mode. The dynamic range of our data was 10 bit (range 0–1024). This limitation is due to the capabilities of the used CTA. Informed consent was obtained from all subjects involved in the study (all have agreed to disclose the following data).

In our paper, we work with real brain aneurysm data from the doctors who annotated the data (the data were annotated by experts). Annotated data are a crucial component for training and validating machine learning models, especially in the field of medical image analysis. We look at a patient’s aneurysm from different angles. We have several different views of the aneurysm from one patient (and several aneurysms from one patient). The neural network learns from different data. The images from the same patient have no similarities between classes. The specific strategic approaches are:

• Data augmentation: Generate additional training samples by applying various transformations to the existing images, such as rotation, scaling, flipping, and changes in brightness or contrast.
• Domain-specific data acquisition: We collaborate with other institutions or research groups to obtain additional brain aneurysm data.
• Active learning: We iteratively select and annotate samples about which the model is uncertain, gradually expanding the labelled dataset.

It is important to note that the effectiveness of these strategies may vary depending on the specific characteristics of our dataset and the nature of the task.

Figure 4 shows a sample of selected images from four patients. The patient number is the anonymized identifier in our database. The image number is the position in the sequence. Each represents one angle-position X-ray image. It shows scans of one patient per row. Several color mappings we were using improved the visualization of the grayscale data. The bottom row is a unique case where adaptively thresholding obtains the binary image of the vascular system. The computer displays grayscale images only at a depth of eight bits, since the human eye has limits in recognizing deep resolutions. Since the data use a 10-bit depth, we can extend the grayscale image in the false color mapping. Such a step improves the observer’s ability to see more details.

3.1. Data Annotations

The annotation (see Figure 5) represents the rectangular area in the image. It has the upper left corner and lower right corner of the region. Because the patients in our database had no more than one aneurysm, each image has a maximum 0–1 annotation. Only the visible aneurysms have an assigned annotation. Some aneurysms are not visible under a specific angle. In such cases, aneurysms do not have an annotation. In the context of the whole image, the aneurysm is comfortably visible to human beings only in zoomed view.

3.2. Data Extraction

We used the annotations for aneurysm extraction. We chose a size of 64 × 64 pixels concerning the maximum dimensions of the annotations. A total of 1230 aneurysm images were annotated and cropped. For each aneurysm annotation, we cut off 10 vessel slices of non-aneurysm vessels around an aneurysm. We were careful not to overlap the sections with the annotation. We obtained an image database with 1230 aneurysm and 12,300 non-aneurysm image slices.

All cropped images pass by zero-mean normalization. Thus, the original 10-bit grayscale pixel was converted from an original range of 0–1023 to a unit value and shifted by the mean value to a final range from −0.5 to +0.5. Figure 6 illustrates the procedure we used to cut the data for aneurysms and non-aneurysms from the original image. The red rectangle represents the original annotation. For the purpose of training the model, it is inconvenient for us that the annotation region has different dimensionality for different images, patients, and aneurysm sizes. For training, we used a square region centered by the original annotation. It is shown by the green square. Around the annotation, we clipped non-aneurysm vasculars. If a random cut interfered with the annotated region, we ignored it and replaced it with another cut. Because there are disproportionately more healthy vessels, we limited the number of non-aneurysm clippings to 10.

We divided the dataset into training, validation, and test sets.

Table 1. Data partitioning into training, validating, and testing.

Set	Aneurysm	Non-Aneurysm
Training	779	9850
Validating	195	1970
Testing	256	2450
Total	1230	12,300

specifies the numbers of data used in training and testing.

4. Proposed Method

In modeling, we focused on two types of models (proposed 2DCNN model and proposed 3DCNN-LSTM). In convolutional modeling, the original two-dimensional image slices enter the system. On the other hand, in 3DCNN-LSTM convolutional modeling, the original three-dimensional image slices enter the system. Both models have a grayscale input image with a resolution of 64 × 64 pixels. The modeling input already includes zero-mean-normalized data. The output is binary information on whether there is an aneurysm or normal blood vessels in the area.

4.1. Proposed 2DCNN Architecture

The proposed 2D convolutional neural network (2DCNN) architecture (see Figure 7) is designed to classify grayscale images with 235 zero-mean-normalized slices into two classes. The mesh consists of two hidden convolution 2D layers followed by MaxPooling 2D layers. Subsequently, the data are transforming by a Flatten layer and input to Dense layers. The input to the convolution model is a grayscale image with zero-mean-normalized slices. The input shape dimension is 64 × 64. The model begins with two hidden convolution 2D layers that extract features from the input images. Each convolution layer applies a set of filters to the input image, resulting in a set of feature maps. The filters are learned during training to identify relevant patterns in the data. The size of the filters is typically smaller than the input image, and the filters slide over the input image to produce the feature maps.

After the convolution layers, MaxPooling 2D layers are applied to reduce the spatial dimension of the feature maps while preserving the most important features. MaxPooling 2D layers take the maximum value in each small region of the feature maps and output a new set of reduced feature maps. This reduces the number of parameters in the model, allowing for faster computation and reducing the risk of overfitting. The output of the MaxPooling 2D layers is then transformed by a Flatten layer, which converts the 2D feature maps into a 1D feature vector.

This pseudo-code provides a high-level overview of the proposed 2DCNN architecture (Algorithm 1):

Algorithm 1 Pseudo-code of the proposed 2DCNN architecture

1: $tensorflow.keras.models$ 2: $tensorflow.keras.layers$ 3: $X\_train,X\_test,Y\_train,Y\_test=train\_test\_split(X,Y)$ 4: $X\_val,X\_test,Y\_val,Y\_test=train\_test\_split(X\_test,Y\_test)$ 5: $model.add\left(Conv2D\right(filters=16,kernel\_size=(3,3)\left)\right)$ 6: $model.add\left(BatchNormalization\right(\left)\right)$ 7: $model.add\left(MaxPooling3D\right(pool\_size=(2,2)\left)\right)$ 8: $model.add\left(Conv2D\right(filters=16,kernel\_size=(3,3)\left)\right)$ 9: $model.add\left(BatchNormalization\right(\left)\right)$ 10: $model.add\left(MaxPooling3D\right(pool\_size=(2,2)\left)\right)$ 11: $model.add\left(Flatten\right(\left)\right)$ 12: $model.add\left(Dense\right(32\left)\right)$ 13: $model.add\left(Dropout\right(0.3\left)\right)$ 14: $model.add\left(Dense\right(2\left)\right)$

The flattened feature vector is passed through a series of Dense layers that perform classification. The Dense layers are fully connected layers that take input from all neurons in the previous layer and output a new set of features. Finally, like the sequential model, there is a binary classification into two classes with SoftMax activation. The SoftMax activation function normalizes the output to produce probabilities for each class, and the predicted class is the one with the highest probability. Further details of the layers are shown in

Table 2. The description of 2D convolution neural network model.

Layer Type	Output Shape	Param #	Kernel Size	Pooling Size	Activation
Conv 2D	62, 62, 16	160	3		ReLU
MaxPooling2D	31, 31, 16			2, 2
Conv 2D	29, 29, 16	2320	3		ReLU
MaxPooling2D	14, 14, 16			2, 2
Flatten	3136
Dense	32	100,384			ReLU
DropOut	rate = 0.3
Dense	2				SoftMax

.

4.2. Proposed 3DCNN-ConvLSTM Architecture

The 3DCNN-based approaches, leveraging 3D convolution layers, have garnered significant popularity for feature extraction from input image data in the biomedical field. Accordingly, our proposed 3DCNN architecture integrates a three-dimensional (3D) convolution layer. This layer functions by employing a sliding 3D convolution window across the input data, positioned above the data itself. Within this window, multiple filters are deployed, with each filter specializing in detecting distinct patterns. These 3D filters traverse the input data in all three spatial directions. The novel neural network architecture integrates 3DCNN (Conv3D) layers, ConvLSTM layer and 2DCNN (Conv2D) layer (see Figure 8). This architecture, denoted as 3DCNN-ConvLSTM, comprises several 3DCNN (Conv3D) layers, succeeded by a solitary ConvLSTM layer, and 2DCNN layer (Conv2D). The description of the layers of the proposed 3DCNN-ConvLSTM can be seen in

Table 3. The description of 3DCNN-ConvLSTM neural network model.

Layer Type	Output Shape	Param #	Kernel Size	Pooling Size	Activation
Conv 3D	62, 62, 62, 64	8704	3		ReLU
MaxPooling3D	31, 31, 31, 64	0		2, 2, 2
Batch Norm	31, 31, 31, 64	256
Conv 3D	32, 47, 46, 64	101,757	3		ReLU
MaxPooling3D	16, 23, 23, 64	0		2, 2, 2
Batch Norm	16, 23, 23, 64	256
Conv 3D	12, 19, 19, 128	457,482	3		ReLU
MaxPooling3D	6, 9, 9, 128	0		2, 2, 2
Batch Norm	6, 9, 9, 128	512
Conv 3D	1, 9, 9, 256	197,855	3		ReLU
Conv 3D	1, 9, 9, 512	132,758	3		ReLU
MaxPooling3D	1, 9, 9, 512	0		2, 2, 2
Batch Norm	1, 9, 9, 512	2048
ConvLSTM	8, 8, 64	1,328,375	3
Batch Norm	8, 8, 64	256
Conv 2D	6, 6, 16	9235	3		ReLU
Flatten	20,736
DropOut	rate = 0.3
Dense	2				SoftMax

. This architecture is divided into the following layers:

• 3DCNN layers (Conv3D) improve the recognition of 3D and dynamic images. Each layer contains a 3D filter that scans data in three directions (x, y, z). The 3D convolution process generates a convolutional map, crucial for thorough data analysis while also preserving essential time and volumetric context information.
• MaxPooling layers, specifically MaxPooling3D designed for 3D data, play a crucial role in diminishing the dimensions of image data. MaxPooling3D is a mathematical operation crafted for handling 3D data, encompassing spatial or spatiotemporal data. These layers are configured using n × n × n regions as filters to execute the max pooling operations. Additionally, a stride parameter is specified, determining the number of pixels the filter advances with each step as it traverses the image.
• The ConvLSTM layer is responsible for handling the features extracted by the Conv3D layers and effectively capturing the temporal dependencies that exist among the frames.
• Batch normalization (Batch Norm) is a technique employed to standardize the output of the preceding layer for each batch of data.
• The Dense layer consists primarily of fully connected neurons. The Flatten layer facilitates the transformation of the matrix into an output vector.

The architecture is predominantly composed of 3D convolution layers but also incorporates essential layers such as flatten and dense. Certain hyperparameters, including the number of filters and the core size of 3D convolution layers and MaxPooling, are constrained by mathematical operations to ensure non-negativity and integer output from the layers. Building upon our previous research [22,23], we introduced an optimization algorithm for ConvLSTM, which was slightly modified and applied to this network. The suggested architecture harnesses the advantages of both Conv3D and ConvLSTM networks. This design includes multiple 3D convolutional layers, one ConvLSTM layer, and one 2D convolutional layer, accompanied by batch normalization, a flattening layer, and a dense layer.

This pseudo-code provides a high-level overview of the proposed 3DCNN-ConvLSTM architecture (Algorithm 2):

Algorithm 2 Pseudo-code of the proposed 3DCNN-ConvLSTM architecture

1: $tensorflow.keras.models$ 2: $tensorflow.keras.layers$ 3: $X\_train,X\_test,Y\_train,Y\_test=train\_test\_split(X,Y)$ 4: $X\_val,X\_test,Y\_val,Y\_test=train\_test\_split(X\_test,Y\_test)$ 5: $model.add\left(Conv3D\right(filters=64,kernel\_size=(3,3,3)\left)\right)$ 6: $model.add\left(BatchNormalization\right(\left)\right)$ 7: $model.add\left(MaxPooling3D\right(pool\_size=(2,2,2)\left)\right)$ 8: $model.add\left(Conv3D\right(filters=128,kernel\_size=(3,3,3)\left)\right)$ 9: $model.add\left(BatchNormalization\right(\left)\right)$ 10: $model.add\left(MaxPooling3D\right(pool\_size=(2,2,2)\left)\right)$ 11: $model.add\left(Conv3D\right(filters=256,kernel\_size=(3,3,3)\left)\right)$ 12: $model.add\left(BatchNormalization\right(\left)\right)$ 13: $model.add\left(MaxPooling3D\right(pool\_size=(2,2,2)\left)\right)$ 14: $model.add\left(Conv3D\right(filters=512,kernel\_size=(3,3,3)\left)\right)$ 15: $model.add\left(BatchNormalization\right(\left)\right)$ 16: $model.add\left(MaxPooling3D\right(pool\_size=(2,2,2)\left)\right)$ 17: $model.add\left(ConvLSTM\right(filters=64,kernel\_size=(3,3,3)\left)\right)$ 18: $model.add\left(BatchNormalization\right(\left)\right)$ 19: $model.add\left(Conv2D\right(filters=16,kernel\_size=(3,3,3)\left)\right)$ 20: $model.add\left(Flatten\right(\left)\right)$ 21: $model.add\left(Dropout\right(0.3\left)\right)$ 22: $model.add\left(Dense\right(2\left)\right)$

The values in

Table 4. The overall number of parameters of proposed 3DCNN-ConvLSTM architecture.

Parameters of the Proposed 3DCNN-LSTM Architecture	Number of Parameters
Total parameters	2,457,427
Trainable parameters	2,527,820
Non-Trainable parameters	1789

describe the overall number of parameters as well as the number of trainable and untrainable parameters.

The proposed 3DCNN-ConvLSTM architecture in this study comprises six 3D convolutional layers and four MaxPooling3D layers. The number of filters and their sizes vary across these layers. The input images have dimensions of 32 × 32 × 3 (width, height, number of channels). The initial 3D convolution layer utilizes 64 filters with a kernel size of 3 × 3 × 3. Following this layer, the MaxPooling3D layer of size 2 × 2 × 2 with a stride of 2 is applied to reduce the data dimension. The second 3D convolution layer maintains 64 filters with the same 3 × 3 × 3 filter size as the first layer. Another MaxPooling3D layer, identical in size to the previous one, follows this layer. The third and fourth 3D convolution layers feature 128 filters, each with a size of 3 × 3 × 3. After the fourth 3D convolution layer, a MaxPooling3D layer, identical in size to the previous ones, is applied. The final two 3D convolution layers are equipped with 256 filters of size 6 × 1 × 1 and 512 filters of size 1 × 1 × 1, respectively. Following these layers, a final MaxPooling3D layer of size 2 × 2 × 2 is applied. Each MaxPooling3D layer is accompanied by batch normalization. At the network’s conclusion, both a dense layer and a flattened layer are present. The dense layer is singular and directly yields an output of 2. Filter sizes were carefully chosen to ensure integer output consistency. The optimization algorithm employed in the architecture is “Adam”, with a default learning rate of 0.1.

5. Evaluation and Results

In this section, we will discuss the experimental results achieved for aneurysm detection using a classification model. The classification model maps predicted classes with labeled instances, and recognition labels are used to distinguish between the actual class and predicted class. The experimental results for aneurysm detection are based on a dataset that includes medical images of the brain. The images are labeled as either containing an aneurysm or not. The classification model is trained using a subset of the dataset, and the performance of the trained model is evaluated using another subset of the dataset. To evaluate the performance of the model, several metrics are used, including accuracy, sensitivity, specificity, precision, and F1 score. Accuracy is the percentage of correctly classified instances, sensitivity is the percentage of true positives (aneurysms correctly identified), specificity is the percentage of true negatives (non-aneurysms correctly identified), precision is the percentage of true positives out of all predicted positives, and F1 score is the harmonic mean of precision and recall. All experiments were executed on a 13th Gen Intel(R) Core(TM) i5-13400F with a clock speed of 2.50 GHz and 32.0 GB RAM. The achieved results were obtained using Nvidia CUDA libraries, and the experiments were conducted on hardware featuring an Nvidia graphics card (NVIDIA GeForce RTX 3060 graphics card), an optimal choice for deep learning models. This graphics card is equipped with advanced hardware support specifically designed for deep learning computations.

The outcomes of the classification model are illustrated through a confusion matrix, detailing the counts of true positives, false positives, true negatives, and false negatives. True positives represent correctly identified aneurysms, false positives denote non-aneurysms incorrectly identified as aneurysms, true negatives indicate correctly identified non-aneurysms, and false negatives signify aneurysms incorrectly identified as non-aneurysms. In a classification model, the mapping of predicted classes with labeled instances is essential. Recognition labels, assigned to classification results generated by the model, serve to distinguish between the actual class and the predicted class. This process yields four possible outcomes:

• True Positive (TP)—The instance is positive and correctly classified as positive.
• True Negative (TN)—The instance is negative and correctly classified as negative.
• False Positive (FP)—The instance is negative but incorrectly classified as positive.
• False Negative (FN)—The instance is positive but incorrectly classified as negative.

True positives (TP) denote cases where the model accurately predicted the presence of an aneurysm in a patient. False positives (FP) represent cases where the model predicted the presence of an aneurysm, but in reality, the patient did not have one. False negatives (FN) indicate cases where the model predicted the absence of an aneurysm, but the patient did indeed have one. True negatives (TN) are instances where the model correctly predicted the absence of an aneurysm in a patient. The confusion matrix serves as a foundation for calculating performance metrics for the model, including accuracy, precision, recall, and the F1 score. In this binary classification scenario, these parameters can be derived directly from the confusion matrix. Various metrics are available for evaluation, with specific emphasis on Specificity (true negative rate (TNR), Sensitivity (true positive rate (TPR)/Recall (R)), positive predictive value (PPV/Precision (P)), and negative predictive value (NPV). Specificity is defined as the probability of a negative test result given that the individual is truly negative:

$$TNR=\frac{TN}{FN+FP}.$$

(1)

Sensitivity (also known as recall) is the probability of a positive test result, conditioned on the individual truly being positive:

$$TPR=R=\frac{TP}{TP+FN}.$$

(2)

The positive predictive value (also known as precision) is the proportion of positive results in statistic that are true positive:

$$PPV=P=\frac{TP}{TP+FP}.$$

(3)

The negative predictive value is the proportion of negative results in statistic that are negative:

$$NPV=\frac{TN}{TN+FN}.$$

(4)

F1 score is a machine learning evaluation metric that combines precision and recall scores, also known as F1-measure, or F1-score (it is used as a measure for model accuracy):

$$F1=2\times \frac{P.R}{P+R}=2\times \frac{PPV.TPR}{PPV+TPR}.$$

(5)

Accuracy is computed as

$$ACC=\frac{TP+TN}{TP+FN+FP+TN}.$$

(6)

Using these metrics, it is possible to evaluate the performance of the classifier and identify areas where it may be struggling. For example, if the classifier has low precision, it may be making too many false positive predictions. If the classifier has low recall, it may be missing too many actual positive examples.

The training accuracy is the accuracy of the model on the training dataset. It measures how well the model is able to fit the training data and is computed by comparing the predicted outputs of the model to the true outputs of the training data. Validation accuracy, on the other hand, is the accuracy of the model on a separate validation dataset. The validation dataset is used to evaluate the performance of the model on unseen data, which is important to ensure that the model is not overfitting to the training data. During the training phase, the model is trained iteratively on the training dataset and the validation accuracy is computed after each iteration to monitor the model’s performance on unseen data. The goal is to optimize the model’s performance on the validation dataset, while avoiding overfitting to the training dataset. In general, the training accuracy tends to increase as the model is trained on more data and more epochs, while the validation accuracy can either increase or plateau and even decrease when the model starts overfitting. Therefore, monitoring both metrics during the training phase is important to ensure that the model is learning effectively and not overfitting to the training data.

Figure 9 shows the training process of the proposed 2DCNN model (blue line and orange line represent train accuracy and validation accuracy, respectively). We can see that at approximately 60 epochs, the model has reached its potential. Figure 10 shows the training process of the proposed 3DCNN-ConvLSTM model. We can see that the proposed 3DCNN-ConvLSTM model needed fewer epochs to reach its optimum. It also had a better ability to classify aneurysms than the 2DCNN model.

Table 5. The confusion matrix of proposed 2D convolutional model (2DCNN).

Observed/Predicted	Aneurysms	Non-Aneurysms
Aneurysms	0.9843	0.0721
Non-aneurysms	0.0157	0.9279

and

Table 6. The confusion matrix of proposed 3DCNN-ConvLSTM model.

Observed/Predicted	Aneurysms	Non-Aneurysms
Aneurysms	0.9919	0.063
Non-aneurysms	0.0081	0.937

record the results of the two proposed models (2DCNN model and 3DCNN-ConvLSTM model). The main diagonal represents the true positive and true negative rates. The values outside the main diagonal represent the false positive and false negative detection error rates, as shown in Table 5 and Table 6. We achieved similar results to Zhou et al. [16]. Zhou et al. used a 3D-CNN, while we only worked with a 2DCNN, as shown in

Table 7. The comparison of results of the proposed methods results with results of other authors.

Evaluation Metrics	Proposed 2D-CNN	Proposed 3DCNN-ConvLSTM	Hou [18] 1D-CNN	Zhou [16] 3D-CNN	Dou [8] 3D-CNN	Nakao [19] CNN/MIP
Accuracy	0.964	0.978	0.958	n/a	0.953	0.941
F1-score	0.965	0.982	0.957	n/a	0.951	0.939
Specificity	0.991	0.997	n/a	0.823	n/a	n/a
Recall	0.940	0.963	0.952	0.866	0.943	0.937
Precision	0.982	0.994	0.962	n/a	0.958	0.943
NPV	0.937	0.965	n/a	n/a	n/a	n/a

.

From these values, we calculated various metrics for model evaluation and entered the results into Table 7. The first two columns represent the results of our proposed methods. The following columns represent the results of different authors who have solved similar aneurysm detection problems using several DNN models.

From the results, we see that our 2DCNN model achieves similar results to Hou [18]. Slightly better results were given by the 3DCNN-ConvLSTM model, which achieved the highest accuracy despite its simpler structure. However, for a better comparison of the results, we would need to test the models on a uniform database, with a larger number of aneurysms and a greater variety of patients.

6. Discussion

The integration of 3D convolutional neural networks (3DCNNs) with Convolutional Long Short-Term Memory networks (ConvLSTM) in our approach has proven to be a pivotal element. This combination allows for a comprehensive analysis of the volumetric data, capturing both spatial and temporal features. This is particularly crucial in the context of brain aneurysm detection, where the dynamic nature of blood flow and the evolution of vascular structures over time play a significant role. One of the key advantages of our approach is the model’s enhanced sensitivity to temporal changes associated with aneurysms. The ConvLSTM enables the network to capture the temporal context, allowing it to recognize subtle alterations in the vascular structures. This is especially relevant for aneurysms that may exhibit dynamic morphological changes over consecutive frames in medical imaging sequences.

The spatiotemporal feature learning capability of our model contributes to an improved accuracy in aneurysm detection. By considering both spatial and temporal dependencies, the network is better equipped to distinguish between normal vascular variations and potential aneurysmatic anomalies. This leads to a reduction in false positives, enhancing the overall reliability of the detection system. Our novel approach demonstrates promising adaptability to patient-specific variations in aneurysm presentation. The ConvLSTM’s ability to capture long-term dependencies enables the model to learn and adapt to the unique characteristics of aneurysms in individual patients. This contributes to a more personalized and robust detection mechanism across diverse clinical scenarios.

In cases where medical imaging provides four-dimensional (4D) data, our approach effectively utilizes this additional temporal dimension. The 3DCNN + ConvLSTM architecture proves advantageous in extracting meaningful insights from the time-dependent evolution of vascular structures, further enhancing the accuracy of aneurysm detection in dynamic imaging sequences.

Despite the promising results, challenges remain, including the need for large, diverse datasets and addressing potential biases in the training data. Ethical considerations, such as patient privacy and model interpretability, need careful attention as the technology advances. Striking a balance between innovation and ethical responsibility is crucial for the responsible deployment of such models in clinical practice.

7. Conclusions

In this article, we focused on the detection of cerebral aneurysms. We used deep neural networks for this purpose. We compared two types of DNNs, namely 2DCNN and 3DCNN-ConvLSTM. In both cases, we had comparable results in terms of accuracy. The convolutional 2D model slightly improved the scoring on the test set. The 2DCNN model was computationally more demanding but reached its optimum after approximately 60 epochs. Afterwards, it showed signs of overfitting. The 3DCNN-ConvLSTM model reached its limits around 100 epochs. In our work, we would like to develop an annotation tool that would use the proposed model to quickly search regions with potential aneurysms. These detections would then be confirmed or rejected by a specialist radiologist. The new data would refine the existing model. Some metrics were not reported by foreign authors; for this reason, we made the results somewhat redundant to make them easier to compare.

Aneurysm classification is an important task in medical imaging and is crucial for the diagnosis and treatment of brain aneurysms. The traditional method of aneurysm classification relies on manual inspection of medical images by a trained radiologist. However, with the advancements in deep learning and computer vision, automated aneurysm classification using deep neural networks has gained a lot of attention in recent years. In this discussion, we will explore the current state of aneurysm classification using deep neural networks and suggest some future directions for research in this field. Deep neural networks have shown promising results in various medical imaging tasks, including aneurysm classification. These networks can automatically extract relevant features from medical images and classify them into different categories. Convolutional neural networks (CNNs) are a popular type of deep neural network that has been used for aneurysm classification. CNNs have been used to extract features from medical images and to classify them into different categories such as aneurysm size, shape, and location. One of the challenges in aneurysm classification using deep neural networks is the lack of annotated medical images. Annotated medical images are necessary to train deep neural networks. However, the availability of annotated medical images is limited due to various reasons such as privacy concerns and the high cost of annotation. Therefore, researchers have used transfer learning to overcome this challenge. Transfer learning involves training deep neural networks on a large dataset of non-medical images and then fine-tuning the network on a smaller dataset of medical images. This approach has shown promising results in aneurysm classification. Another challenge in aneurysm classification using deep neural networks is the class imbalance problem. In medical imaging, the number of normal cases is usually much larger than the number of abnormal cases. This class imbalance problem can result in a biased model that performs well on normal cases but poorly on abnormal cases. Researchers have used different techniques to address the class imbalance problem, such as oversampling the minority class, undersampling the majority class, and generating synthetic data.

The proposed approach for brain aneurysm detection using deep neural networks holds significant applicability in the field of medical imaging and healthcare. The method leverages deep neural networks to analyze medical imaging data, indicating its applicability in the advancement of diagnostic capabilities within medical imaging technologies. Deep neural networks have shown prowess in learning intricate patterns, enabling the model to accurately detect and identify brain aneurysms in medical images. This contributes to enhanced diagnostic precision. Early detection of brain aneurysms is crucial for timely intervention and treatment planning. This method provides a means to identify potential issues in their nascent stages, facilitating proactive medical measures. Automated detection through deep neural networks can significantly reduce the manual workload of healthcare professionals. By assisting in preliminary analysis, the method allows medical experts to focus on in-depth evaluation and decision making. The use of deep neural networks contributes to the speed and efficiency of the diagnostic process. Rapid identification of brain aneurysms can be pivotal in emergency situations or time-sensitive medical conditions. The approach can be adapted to various imaging modalities, such as magnetic resonance imaging (MRI) or computed tomography (CT) scans, making it versatile for different medical imaging scenarios. The deep neural network aspect implies the potential for continual learning and improvement. As the model encounters more data, it can adapt and refine its detection capabilities, staying current with evolving medical imaging practices. The method can contribute to ongoing research and development efforts in the field of medical image analysis and neural network applications, fostering innovation and advancements in healthcare technologies.

In conclusion, aneurysm classification using deep neural networks has shown promising results in recent years. However, there are still some challenges that need to be addressed, such as the lack of annotated medical images and the class imbalance problem. Future research can focus on developing more robust deep neural networks, exploring the use of multimodal imaging, and developing more comprehensive datasets of annotated medical images.

In our future work, we would like to focus on the semi-automated analysis of CTA images to calculate the probability of aneurysms, where the physician could check for highlighted potential aneurysm areas in the first step. We also wanted to try to compute a 3D model from the images and test the proposed procedures with 3D-CNNs, possibly trying to train the models on already pre-trained networks. One direction for future research is to develop more robust deep neural networks that can handle noise and artifacts in medical images. Another direction is to explore the use of multimodal imaging for aneurysm classification, where multiple types of medical images, such as CT and MRI, are used together to improve the accuracy of the classification. Finally, the development of more comprehensive datasets of annotated medical images can help to improve the performance of deep neural networks in aneurysm classification.