Multi-Layer ResNet-DenseNet architecture in consort with the XgBoost classifier for intracranial hemorrhage (ICH) subtype detection and classification

Abstract

Intracerebral haemorrhage (ICH) is defined as bleeding occurs in the brain and causes vascular abnormality, tumor, venous Infarction, therapeutic anticoagulation, trauma property, and cerebral aneurysm. It is a dangerous disease and increases high mortality rate within the age of 15 to 24. It may be cured by finding what type of ICH is affected in the brain within short period with more accuracy. The previous method did not provide adequate accuracy and increase the computational time. Therefore, in this manuscript Detection and Categorization of Acute Intracranial Hemorrhage (ICH) subtypes using a Multi-Layer DenseNet-ResNet Architecture with Improved Random Forest Classifier (IRF) is proposed to detect the subtypes of ICH with high accuracy, less computational time with maximal speed. Here, the brain CT images are collected from Physionet repository publicly dataset. Then the images are pre-processed to eliminate the noises. After that, the image features are extracted by using multi layer Densely Connected Convolutional Network (DenseNet) combined with Residual Network (ResNet) architecture with multiple Convolutional layers. The sub types of ICH (Epidural Hemorrhage (EDH), Subarachnoid Hemorrhage (SAH), Intracerebral Hemorrhage (ICH), Subdural Hemorrhage (SDH), Intraventricular Hemorrhage (IVH), normal is classified by using Improved Random Forest (IRF) Classifier with high accuracy. The simulation is activated in MATLAB platform. The proposed Multilayer-DenseNet-ResNet-IRF approach attains higher accuracy 23.44%, 31.93%, 42.83%, 41.9% compared with existing approaches, like Detection with classification of intracranial haemorrhage on CT images utilizing new deep-learning algorithm (ICH-DC-CNN), Detection with classification of intracranial haemorrhage on CT images utilizing new deep-learning algorithm (ICH-DC-CNN-ResNet-50), Shallow 3D CNN for detecting acute brain hemorrhage from medical imaging sensors (ICH-DC-S-3D-CNN), Convolutional neural network: a review of models, methods and applications to object detection (ICH-DC-CNN-AlexNet) respectively.

Keywords

Acute Intracranial Hemorrhage (ICH)Computerized Tomography (CT)Residual Network (ResNet)Densely Connected Convolutional Networks (DenseNet)Extreme Gradient Boosting (XGBoost) Classifier

1 Introduction

Intracranial hemorrhage refers to the acute bleeding inside the skull or brain [1]. The subtypes of intracranial hemorrhage are: subarachnoid hemorrhage, Intraparenchymal hemorrhage, subdural hemorrhage, epidural hemorrhage and Intraventricular hemorrhage [2]. Any form of bleeding or its consolidation can make traumatic brain injury, cerebrovascular pathology, arterial hypertension, and even surgical intervention more difficult [3]. Computed tomography is a quick and trustworthy diagnostic tool for acute hemorrhage diagnosis [4]. An automatic hemorrhage identification in CT scans with computer-aided design can helps to rapidly detect the determination on treatment [5, 6]. Even though, the count of neuro imaging data available is commonly limited for the development of these solutions [7, 8]. To diagnose ICH, Computed tomography is a non-invasive imaging tool [9, 10]. The Hemorrhage can be identified in non-contrast CT, because blood has slightly higher density (Hounsfield unit (HU)) over other brain tissues, but lesser than bones [11, 12]. The exact diagnosis of bleeding is very important for physicians to take medical interventions [13, 14]. Moreover, after working hours, the assessment of head CT is essential in the emergency department [15, 16]. At most medical centers, early interpretations of the head CT are generally performed by junior radiologists, radiotherapists, emergency physicians to provide adequate care to patients of medical importance [17 –19]. The early interpretations are expressed later by experienced radiologists [20]. Numerous researches have assured that there are discrepancies amid the first and last interpretations, and that certain misinterpretations may have clinical effects [21]. ICH is reported to have 13.6% (141/1037) discrepancies as well as misidentified ICH general subtypes are SDH and SAH attains 39%, and 33% [22]. An automatic triage scheme to exact ICH identification is desired for lessening the misidentified rate. Nowadays, artificial intelligence is significance in the field of clinical imaging [23]. Among these, certain studies try to detect abnormalities in head CT involving ICH utilizing deep learning/machine learning models [24 –28].

The healthcare sector is quite dissimilar from other fields [29]. It is a higher priority sector; also people expect more care and cost services regardless. After deep learning success in another real-world application, it offers exciting solution through better accuracy in medical imaging as well as key technique for future applications in the health sector [30, 31]. Recognition of automated brain tumor in Magnetic resonance imaging is a complicated task due to size complex, locality, variability. Several methods were presented previously to detect the brain stroke lesions, but no other methods provide sufficient accuracy as well as maximized the error rate. But still, there is a scope to design a suitable procedure for creating and activating useful classification scheme for brain tumour. These draw backs have incited to do this work.

•Innovation

The novelty/innovation of this paper is that the proposed Acute Intracranial Hemorrhage (ICH) subtype detection and classification approach using a multi-layer ResNet-DenseNet architecture in consort with the XGBoost Classifier (ICH-ResNet-DenseNet-XGBoost) algorithm is used to classify the acute intracranial hemorrhage (ICH) disease into EDH, SDH, SAH, ICH and IVH with high classification accuracy. The XGBoost has certain merits, like robust generalization capability, higher expandability, rapid computing speed. This is composed of series of base classifiers: decision tree, KNN, SVM, logistic regression and more. After computing the base classifier, the base classifiers are superimposed linearly to optimize the algorithm.

•Motivation

The main aim of this work is to enhance the classification accuracy by lessening the computational time. The CT images of head are utilized in a great extent to diagnose many neurological diseases like acute traumatic brain injury (TBI), acute stroke and ICH. In this study, one of the most crucial neurological abnormalities namely, acute intracranial hemorrhage is detected and classified into various subtypes for easy identification and early diagnosis. Since, early diagnosis is the only way to control the mortality rate in patients suffering from this deadly disease.

The main contributions of this work are summarized as follows:

•An automatic detection with classification framework for recognizing and categorizing subtypes of acute intracranial hemorrhage (ICH) in the brain is presented.

•At first, preprocessed using the Altered Phase Preserving Dynamic Range Compression for eliminating the unnecessary noise and outliers from the raw CT images.

•Then, a ResNet combined with DenseNet with multiple Convolutional layers is adopted to extract the critical features from the head CT images.

•Finally, an extreme gradient boosting (XGBoost) algorithm is utilized to classify the subtypes of ICHs with high classification accuracy.

•XGBoost has certain merits, like robust generalization capability, higher expandability, rapid computing speed.

•The training and validation is done by employing head CT imageries taken from a Physionet repository publicly available ICH CT image dataset.

•The effectiveness of the proposed model is evaluated based on several performances evaluating metrics, viz accuracy, sensitivity, specificity, ROC curve, misclassified error.

•The experimental outcomes obtained from the proposed model are compared with existing method such as are illustrated below in a tabular form. Here the proposed method ICH-DC-ResNet-DenseNet-XGBoost compared with existing approaches, such as ICH-DC-CNN [40], ICH-DC-CNN-ResNet-50 [41], ICH-DC-S-3D-CNN [42], ICH-DC-CNN-AlexNet [43].

The rest of this paper is mentioned as below: section 2 presents recent literature works, section 3 illustrates proposed methodology, section 4 exemplifies the experimental outcomes, section 5 presents the conclusion.

2 Related works

Among various investigation works related to detection and classification of ICH hemorrhages, some of the most recent research works are reviewed as follows:

Anupama et al., [32] have presented the deep learning (DL)–base ICH identification using Grab Cut-base segmentation along Synergic DL (SDL) termed GC-SDL. The presented model develops Gabor filtering for the removal of noise and to raise the quality of image. Grab Cut-base segmentation was employed to effectively detect the diseased parts via the imagery. SDL was used to activate the process of feature extraction, finally, softmax layer was deemed as the classifier. To investigate the GC-SDL presentation, a set of testing taken by benchmark ICH dataset, then the outcomes were examined utilizing various evaluation metrics. The outcomes show that the GC-SDL model attained 94.01% sensitivity, 97.78%, specificity, 95.79% precision, 95.73% accuracy. The limitation of this paper was the accuracy is low.

Zhang et al., [33] have presented a model for detect artificial lesions in non-lesion CT imageries. Artificial masks of size, shape were created utilizing Artificial Mask Generator and converts hemorrhage lesions with the help of Lesion Synthesis Network. Imageries with and without artificial lesions were consolidated for training an ICH detection including Residual Score. Then examine the method by the auxiliary diagnosis task of ICH. The outcomes display the presented approach reach AUC value 84% to 91% for ICH detection, 89% to 96% for classification task. In this time was increased.

Karki et al., [34] have presented a remotely supervised model for automatically computing the best window settings by combining window estimator module (WEM) with deep convolutional neural network base conjunction classifier. Apart from predicting feasible window settings, first 4 window settings statistically calculate mean and standard deviations for whole datasets. Experimental results show that utilizing the first 4 window settings recognized through window evaluator module were optimally efficient. In this error rate was increased.

Mansour et al., [35] have presented an Artificial Intelligence with Big Data Analytics-Base ICH E-Diagnosis (AIBDA-ICH) utilizing CT imageries. The method uses IoMT devices to the process of data acquisition. AIBDA-ICH method includes a cut-base segmentation mode to identify suffered areas in CT imageries. To manage large data, the Hadoop Ecosystem with its modules was employed. To derive proficient set of feature vectors, the CapsNet was used as a feature extractor. The presented method makes the process of classifying the fuzzy DNN model. To examine the performance of the AIBDA-ICH, simulation was done and the results were examined under various aspects. The outcomes shows 94.96% and 98.59% accuracy and precision, respectively. In this accuracy was decreased and time was increased.

Hausman-Kedem et al., [36] have presented Monogenic Causes of Apparently Idiopathic Perinatal Intracranial Hemorrhage. The whole-exome sequencing (WES) uses pICH and their parents to differentiate between fetuses and neonates. The cause of the variance was deemed as per American College of Medical Genetics and Genomics criterion, the consistency amid the recommended genes and phenotype with inheritance pattern. 26 probands were added. Intraventricular hemorrhage was the typical type of hemorrhage (n = 16, 62%), that was emulated by subpial (n = 4, 15%), subdural (n = 4, 15%), parenchymal (n = 2, 8%) hemorrhage. The presented method does not support large dataset and accuracy was decreased.

Ye et al., [37] have suggested a three-dimensional joint convolutional and recurrent neural network (CNN-RNN) to recognize intracranial hemorrhage (ICH) and its 5 subtypes. A total of 2836 subjects were included from the 3 institutions, and overall 76,621 slices were added from the Contrast Head CT scans in a retrospective method approved by this process. Where, 90% of the data was taken for training, remaining 10% for testing. To implement 3D CT scan, that algorithm was took less than 30 seconds on average. The limitation of this paper was the computational time is increased and the speed is reduced.

Patel et al., [38] have presented Image Level Training with Prediction: Intracranial Hemorrhage recognization at 3D Non-Contrast CT. The method was was presented for the ICH in 3D non-contrast CT. CNN trained to identify ICH on axial pieces. LSTM was examining sequential information derived through slice level classifications. The presented model was trained from end-end utilizing entire non-higher resolution 3D CTs. Overall 1554 cranial CTs were applied for training, 386 images for testing. The outcomes displayed an area below the receiver operating characteristic curve was 0.96. The average time was approximately 0.5 seconds for classification is. In this error rate was increased.

Hu et al., [39] have presented ICH segments DNN for the automatic separation of regions. First, it introduced an encoder-decoder CNN (ED-Net) lower and higher level semantic information. Second, it introduces were synthetic loss function by focusing on smaller ICH regions to deal with the issue of data imbalance. Third, it improves the clinical adaptability to collect 480 patients from 4 hospitals to create a multi-center database. Finally, ED-Net was evaluated in a multi-center ICH clinical database from various parameters with loss functions. In this computational time was decreased.

Lee et al., [40] have presented a novel deep-learning algorithm for artificial neural networks (ANNs), it was separated from the back-propagation method. The aim was to analyze the flexibility of using the approach for intracranial haemorrhage detection with classification of its subtypes, without utilizing CNN. To intracranial haemorrhage identification with the summation every computed tomography (imageries for each case, the area under the ROC curve was higher, and the sensitivity was lower.

Sage and Badura [41] have presented a model to recognize the category of intracranial hemorrhage types in head computed tomography imageries. The model trained for every hemorrhage subtype was depending on dual branch CNN of ResNet-50. It extracts features from 2 input datas: image concatenation normalized in dissimilar intensity windows, stack of 3 consecutive slices making 3D spatial context. It provides lower computational time with higher f-measure.

Singh et al., [42] have presented a model for normalizing 3D volumetric scans utilizing intensity profile of training samples. It assists the CNN by making maximal contrast surround the abnormal region of interest in the scan. Then utilize CQ500 head CT dataset to prove the validity of presented method for identifying various acute brain hemorrhages, like subarachnoid hemorrhage, intraparenchymal hemorrhage, subdural hematoma, Intraventricular hemorrhage. It provides higher error rate with lower precision.

Dhillon and Verma [43] have presented a detailed review of various deep architectures emphasizing specific model properties. The CNN architectures functioning of was described, its components emulated by classical LeNet model to AlexNet, ZFNet, GoogleNet, VGGNet, ResNet, ResNeXt, SENet, DenseNet, Xception, PNAS/ENAS. The deep learning architectures focused on 3 applications (i) wild animal identification, (ii) small arm identification (iii) human being identification. It provides lower computational time with higher recall.

2.1 Motivation behind this research work

Deep learning techniques are becoming popular due to its ability to extract fine information from medical imageries. Radiological imaging techniques, viz CT, MRI, comprise multiple imageries volumetric stacks than individual imageries. The 3-D nature of these exams creates difficulties during ICH recognition. Therefore, precise segmentation and detection is considered the most essential objective. Also, the location of abnormalities on CT of the head is also significant in the decision-making process, as clinicians will forever require personally visualizing and confirming the abnormalities location on CT examination of the head [44]. Though, manual analysis executed through radiologists is difficult with time consume. Such circumstances, rapid, automated detection with classification of ICH are utmost significance.

3 Proposed methodology

The CT images of head are utilized in a great extent to diagnose many neurological diseases like acute traumatic brain injury (TBI), acute stroke and ICH. In this paper, one of the most crucial neurological abnormalities namely, acute intracranial hemorrhage is detected and classified into various subtypes for easy identification and early diagnosis. Since, early diagnosis is the only way to control the mortality rate in patients suffering from this deadly disease.

3.1 Overall diagram of proposed acute ICH detection and classification framework

The overall diagram of the proposed acute ICH detection and classification framework is represented below on Fig. 1 as follows:

Fig. 1

Overall workflow of Proposed Framework.

3.2 Pre-processing using Altered Phase Preserving Dynamic Range Compression (APPDRC)

Pre-processing is the first step in eliminating noise, improving the quality of the input, and identifying the object of interest. There are various challenges in automating the processing of medical images. These images are often affected by noise blurriness, and can have less contrast, resulting in difficulty in observation and detecting diseases. Medical images vary based on medical imaging system parameters and the brightness, contrast, sharpness, and visibility of structures across manufacturers. Therefore, the first and most important step is to enhance and standardize the image. In the preprocessing stage, the unwanted noises are filtered using APPDRC method. This method automatically detects the ICH diseases at an early stage. The data are reconstructed with the help of 2D equivalent function of Hilbert transform and resized filter with the frequency domain is f₁, f₂ and it is expressed in Equation (1) and (2) as follows, $S_{1} = w \frac{f_{1}}{\sqrt{f_{1}^{2} + f_{2}^{2}}}$ (1) $S_{2} = w \frac{f_{2}}{\sqrt{f_{1}^{2} + f_{2}^{2}}}$ (2) where the spatial representations of the vector are expressed as S = (S₁, S₂) this is known as the convolutional kernel resize transform and w represents the quantity of the image. The preprocessed images are provided to feature selection process.

3.3 Feature extraction using combined ResNet and DenseNet Framework

The most relevant features from region of interest (ROI) CT image are extracted using ResNet-DenseNet model. The image features are extracted to avoid the over fitting problems during classification process. By this can identify what type of haemorrhages is affected and it can classify the images with great accuracy. Then the parameter explanations of the ResNet-DenseNet Architecture is given below,

3.3.1 ResNet Architecture

For image classification, the ResNet architecture is greatly modularized network structure. This is incited through VGG, Inception, ResNet. ResNet contains residual blocks that shares the similar topology each other. ResNeXt-101 32×8 d designs is same as ResNet-101 that contains 101 layers. The selected model has a cardinality of 32 with bottleneck width of 8. The cardinality size of the set of transformations (alternative neural pathways) is significant. Every ResNet block has 3 convolutional layers, the 1st and 3rd layers containing filters along 1×1 spatial support, 2nd layer has spatial support of 3×3. The count of filters at 2nd convolutional layer is tiny; this layer performs as a bottleneck layer. The bottleneck width implies count of filters in the bottleneck layer. The amount of learnable parameters approximates 88 million, creating 16B FLOPS complexity. ResNet is pre-trained in a weakly supervised manner, the singular CT slices to fine-tune predicting the presence and the category of intracranial hemorrhage per CT slice. The ResNet is defined as the residual neural network (ResNet), which consists of residual learning features, is formulated in Equation (3), $b = R (a, {V_{j}}) + a$ (3) where a and b is represented as the input and the output layers in the network, V_j is represented as the weight in the weight matrix, and R is represented as the residual mapping functions in the learning process. Then the residual mapping functions are given in Equation (4) as follows, $R_{fm} = V_{2} r (V_{1} a + y_{1}) + y_{2}$ (4) $r (a) = max (0, a)$ (5) where r is represented as the rectified Linear unit in the activation functions, V₁, V₂, y₁, y₂ is represented as the weights and the bias in the first and the second layer.

3.3.2 DenseNet Architecture

DenseNet procedure is the one of the type of the CNN architecture. The main function of the DenseNet procedure is that the layers are connected in dense manner. It is the process of every layer in the network is directly linked to the all the blocks. Due to these properties of the dense layers the image features are extracted and can be reused and gives strengthen to the feature transmission. Let us consider the densely connected layers as D_k and the output of the network is defined as the a_k-1 with the layer k - 1^th, D_k (a) represents the series of connections with one layer to the another, then the output k^thlayer with the a_k is formulated in Equation (6) as follows, $a_{k} = D_{k} ([a_{0}, a_{1}, . . . . . a_{k - 1}])$ (6) where [a₀, a₁, . . . . . a_k-1] is represented as the concatenation operation. Assume that the every functions in the D_k will generates the L features, concurrently the number of feature maps in k^thlayer is changed as l₀ + l × (k - 1), Where l₀ implies number of channels exist in input layer. Here, the parameter l is the growth rate and it will maintain the number of parameters in the architecture. The main features of the DenseNet are that it consists of less number of parameters when comparing with the other network and it does not use the learning redundant features. In the proposed framework, a combined architecture formed by combining Res Net using DenseNet is utilized to extract the vital features from the CT image. The combined feature extraction framework utilizes multiple convolutional layers for training the input parameters.

The deployment of large number of convolutional layers can increase the training errors as well as degrade the classification accuracy. These limitations can be solved by employing a combined (ResNet-DenseNet) framework. The combined feature extraction framework uses multiple convolutional layers to learn the input and output parameters. Also, the combination of ResNet with DenseNet improves the training speed and detection accuracy of the proposed deep learning based feature extraction framework.

The ResNet-DenseNet unit is composed of various convolutional layers with ReLU (Rectified Linear Unit) activation function.

During training process, the back-propagation of the gradient is carried out. The process of back-propagation helps to achieve better training in the deep convolutional neural network and enables continuous transmission of information.

During training process, the area is too small, because the resolutions of the convolutional layer are getting degraded. Therefore, to protect the characteristics of small area or low dimension in the proposed study, a dense convolutional network called DenseNet is employed. The DenseNet connects an individual layer with other layers in the network in a feed-forward manner. Here, the dense layer is subsequent to transition layer. Transition layer is composed of 1 × 1 convolutional kernel. The count of channels at the transition layer represents 4k, where the parameter k represents that growth rate. Here, k = 4 implies output feature image obtained from every layer contains dimension of 4. Also, the DenseNet model improves the feature reuse on one level.

In the combined ResNet-DenseNet model, the feature maps from the entire earlier layers is fed as input to each layers and the feature maps of its own is fed as input to the entire next layers. The combined feature extraction framework improves better propagation of features and enhances feature reuse strategy. Therefore, the low-dimensional features can also be identified effectively with the help of ResNet-DenseNet model.

The residuals in the ResNet-DenseNet model deepen the training process and do not eliminate even the low-dimensional features that are also necessary for the prediction process.

The structure of typical residual-dense block is depicted below in Fig. 2 as follows:

Fig. 2

Typical Representation of ResNet-DenseNet model.

Assume the network consists of L convolutional layers, in which each layer implements a non-linear transform function of H_i (·). Here, the parameter idenotes the i^th layer that obtains the feature maps u₀, u₁, . . . . . . . . , u_i-1 as input from all the previous layers. Thus, the corresponding feature map at i^th layer is expressed as, $u_{i} = H_{i} ([u_{0}, u_{1}, . . . . . . . ., u_{i - 1}])$ (7)

In the above expression, the term [u₀, u₁, . . . . . . . . , u_i-1] defines the feature maps obtaining in cascade manner. Also, the term H_i (·) denotes the non-linear transformation function of 2 consecutive processes: ReLU activation function and 3 * 3 convolution (Conv).

The output at c^th convolutional layer is expressed as follows: $V_{d, c} = H {[V_{d - 1}, V_{d, 1}, . . . . . ., V_{d, c - 1}]$ (8) where the parameters V_d-1 and V_d defines the input and output of d^thunit. Here, the number of feature maps is represented as F₀.

The number of feature maps at end unit should be controlled for better feature extraction. This can be done by employing 1 × 1 convolution in the model. It can be expressed as, $V_{d, LV} = H_{LVV}^{d} {[V_{d - 1}, V_{d, 1}, . . . . . ., V_{d, c}]$ (9) where the term $H_{LVV}^{d}$ indicates 1 × 1 convolution.

Finally, the output from combined ResNet+DenseNet unit is expressed as, $V_{d} = V_{d - 1} + V_{d, LV}$ (10)

In the proposed feature extraction framework, a total of 3 dense blocks layers are used 3BN + ReLU + Conv (3 * 3) layer structures.

The statistical features are extracted by ResNet-DenseNet model. Here, ResNet-DenseNet model denotes (a, b) modules. It is scaled with pixel, its intensity is a with corresponding pixel b, the distance represents c. The features are extracted form the central pixels and indicated in the uniform or non-uniform patterns in sampling circular. From this, extract the numerous radiomic features, the radiomic features are classified as mean, variance, entropy, energy.

Here, 4 features are extracted: mean, variance, entropy and energy. The mean value is defined as the value of average mean pixel on RoI for image brightness; this is expressed as, $M (I_{n}) = \frac{1}{e_{i} \times e_{j}} \sum_{X = 1}^{e_{i}} \sum_{Y = 1}^{e_{j}} O (i, j)$ (11) where i, j is represented as the co-ordinates of input image pixels along center of local region for image, e_i, e_j denotes neighbor pixels grey intensity values, O (i, j) denotes image pixels sampling center. Energy is defined as the values of arithmetic mean. The energy is shown as, $Eng = \sum_{j = 1}^{X_{i}} {([i, j (e_{(i)}, e_{(j)})] - mean)}^{2}$ (12) where, i, j is represented as the co-ordinates of input image pixels along center of local region for image, e_i, e_j denotes neighbor pixels grey intensity values. Entropy is measured as the cube of the variance of the image among two different variables functions is shown in Equation (13) $En = \frac{1}{var {iance}^{3}} \sum_{j = 1}^{X_{i}} {([(e_{(i)}, e_{(j)})] - mean)}^{3}$ (13)

The variance is computed as the scaling of distance amid the 2 or more image features for extracted; this is expressed as, $V = \frac{1}{e_{i} \times e_{j}} \sum_{i = 1}^{e_{i}} \sum_{j = 1}^{e_{j}} {(e (i, j) - mean)}^{2}$ (14)

After extracting these features, it is given to the XGBoost classifier for classifying Acute Intracranial Hemorrhage (ICH) subtype as Epidural Hemorrhage, Subdural Hemorrhage, Subarachnoid Hemorrhage, Intracerebral Hemorrhage, Intraventricular Hemorrhage.

3.4 Classification

In this study, the acute ICH is classified into various subtypes like SDH, EDH, ICH, SAH and Intraventricular Hemorrhage (IVH) using an efficient extreme gradient boosting (XG Boost) classifier.

3.4.1 Classification Using XG Boost Algorithm

An Extreme Gradient Boosting (XGBoost) classification algorithm is one of the most efficient scalable machine learning models with tree-boosting strategy. The XGBoost has certain merits, like robust generalization capability, higher expandability, rapid computing speed. This is composed of series of base classifiers: decision tree, KNN, SVM, logistic regression and more. After computing the base classifier, the base classifiers are superimposed linearly to optimize the algorithm. The objective function with prediction function is structured in supervised learning. The training parameters are employed to lessen the objective function to learn the related parameters, also the prediction process as well as acquired parameters is classified or predicts the unknown sample numerically.

XGBoost possess the capability to solve several machine learning problems with its gradient boosting and regularization property. XGBoost is built using a collection of several decision trees that effectively solves the overfiting problem during classification.

At first, a decision-tree is built with several leaf nodes. Here, each node defines a single feature essential for performing classification. Then, the best splitting point is estimated. After that, assign specific weight values for each new leaf nodes. Finally, the nodes with negative gains are deleted by the process of pruning.

Let the feature set V_i contains the feature vectors V₁, V₂, . . . . . , V_m, here, i = 1, 2, . . . . . . , m. Here, feature significance score IS_i implies every feature that is evaluated to distribute training data over the entire tree.

The feature significance score is given as follows, ${IS}_{i} = {X | X = W_{i} V_{i}}$ (15)

In the above expression, the parameter W_i defines the weight value for every feature and the parameter V_i implies the set of feature.

Here, the classification output at i^th data point is expressed as, $Y_{i}^{'} = \sum_{m = 1}^{m} p_{k}$ (16)

In the above expression, the parameter m indicates that number of decision trees and parameter p_k indicates that classification output from a decision tree.

Then, the loss functions for training the XGBoost model is optimized for achieving the best classification results.

The loss function for classifying acute ICH subtypes is given as, $L = - \frac{1}{N} \sum_{i}^{N} ((Y_{i} log f_{i}) + (1 - Y_{i}) log (1 - f_{i}))$ (17)

The Regularization function for controlling the complexity and for preventing the problem of overfiting during classification is given by, $Ω = Ψ^{T} + \frac{1}{2} λ \sum_{j = 1}^{T} W_{j^{2}}$ (18) where, the parameter Tdenotes count of leaves in the decision tree, parameter W_{j
²}denotes the score on j^thleaf.

Thus, objective function ^′O′ of proposed XGBoost model is articulated as, $O = L + Ω$ (19)

In the above expression, the parameter L represents the loss function that estimates how well the proposed classification model is predictive.

In XGBoost classifier, the mean and standard deviation are used as gradient descents for optimizing that objective function.

The objective function of proposed XGBoost model is articulated as, $\begin{matrix} O (t) = \sum_{i = 1}^{n} [g_{i} p_{t} (X_{i}) + \frac{1}{2} h_{i} p_{t}^{2} (X_{i})] \\ + Ω (p_{t}) \end{matrix}$ (20) where, the parameter X_i represents that feature vector at i^th data point and parameter n denotes the number of iteration performed for reaching the maximum depth.

The classification outputs from the proposed XGBoost classifier is tested under various threshold values for obtaining the best classification outcomes with high accuracy.

4 Results

The proposed approach is done in MATLAB site and its performance is compared with existing approaches. The brain images used for the assessment were taken as the Physionet repository in the generally available ICH CT image dataset [45]. The experimental outcomes obtained as proposed strategy with classification models to verify the performance of the ResNet-DenseNet-XGBoost model.

4.1 Implementation details

Dilated ResNet 38 [46] is a backbone of network, every hyper parameters are created in UCSF-4.4K training set. Cross-entropy loss is optimized utilizing 0.99 momentums. The learning rate is reduced through 0.1 for every 160 epochs. At training dataset, for controlling class unbalance amid the positive and negative cases, 30% of the patch samples are taken from the positive images for every training mini-batch with the positive pixel loss over 3 factors up-weighted. Because multiglass experiments are exploratory naturally, they acted without any balance in the categories of positive class. During training, the backbone as well as pixel prediction branch (1 up-convolution layer) are trained in 10^-3 initial learning rate for 400 epochs. Likewise, for 40 epochs, the patch Classification Branch is trained. The whole model changed well for 30 epochs in 5 × 10^-5 learning rate. At the inference time, two-thirds of the adjacent patches are sampled to each other. The predictions of pixel on every patch maps to the imagery location and averaged to give the last prediction. The maximum patch in the stack classification score layer is deemed as classification score. The sample evaluates every layer within an average of 1 second.

4.2 Dataset description

For training and validation, a publicly available dataset called Physionet repository is employed to examine ICH from CT images. Here, the size of each CT image 650 × 650 pixels. In this study, dataset of 82 CT scans is gathered, including 36 scans for patients diagnosed with intracranial hemorrhage like Intraventricular (IV), Intraparenchymal (IP), Subarachnoid (SA), Epidural (ED) and Subdural (SD). Every CT scan includes 30 slices with 5 mm slice-thickness. The mean and standard deviation of patients’ age are 27.8, 19.5. In which, 46 are males and 36 are females. In the proposed study, 60% of CT slices are employed for training and remaining 40% of CT slices are used as testing set. The feature dimension of input CT image is 168 × 299 × 3 pixels, ROI extraction is 147 × 87 × 3 pixels, SAH is 150 × 84 × 3 pixels, EDH is 150 × 76 × 3 pixels, SDH is 153 × 72 × 3 pixels, IVH is 154 × 84 × 3 pixels, and IPH is 160 × 83 × 3 pixels.

The representation of various kinds of acute intracranial hemorrhage images classified from the CT image by the proposed model is depicted below in Fig. 3:

Fig. 3

Representation of various kinds of acute intracranial hemorrhage images.

4.3 Performance evaluation

The results obtained from the proposed ResNet-DenseNet-XGBoost framework is associatedutilize2 state-of-art existing approaches for revealing effectiveness of proposed strategy in terms of classification accuracy, specificity, sensitivity, mean, standard deviation, misclassified error and receiver operating curve (ROC) performance.

The following are significant rules for calculating performance measurements:

•TP represents True Positive

•TN represents True Negative

•FP represents False Positive

•FN represents False Negative

Sensitivity

This is determined utilizing the following expressions: $Sensitivity = \frac{TP}{TP + FN}$ (21)

Specificity

This is determined using the below expression: $Specificity = \frac{TN}{TN + FP}$ (22)

Accuracy

It is the classification performance corresponding to the total count of classification tests. $Accuracy = \frac{TP + TN}{TP + FN + FP + TN}$ (23)

ROC (Receiver Operating Characteristics) Curve

This is a graph that portrays the classification model performance over the entire classification

thresholds. This curve has 2 parameters: true positive rate, false positive rate.

Misclassified Error

This is estimated during classification using the given expression, $ME = \frac{FP + FN}{TP + FN + FP + TN}$ (24)

Mean

Mean refers to the average value of the intensity of the image. It is calculated using the equation as follows: $Mean, η = \sum_{i = 1}^{m} ip (i)$ (25)

The parameter p implies independent data vectors.

Standard Deviation (SD)

The SD σ refers to the estimation of the mean square deviation of the pixel value from their mean value. It is determined using the formula: $σ = \sqrt{Variance} = {(\sum_{i = 1}^{m} {(i - η)}^{2} p (i))}^{2}$ (26)

Simulation phase: 1

Table 1-8 shows the simulation results obtained from the proposed acute ICH detection and classification model for training-testing % are illustrated below in a tabular form. Here the proposed method ICH-DC-ResNet-DenseNet-XGBoost compared with existing approach such as are illustrated below in a tabular form. Here the proposed method ICH-DC-ResNet-DenseNet-XGBoost compared with existing approaches, such as ICH-DC-CNN [40], ICH-DC-CNN-ResNet-50 [41], ICH-DC-S-3D-CNN [42], ICH-DC-CNN-AlexNet [43].

Table 1 depicts the accuracy analysis, the performance of the proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 26.92%, 15.11%, 20.14%, and 30.26% higher accuracy for Epidural Hemorrhage (EDH); 57.14%, 15.47%, 17.68% and 25.64% higher accuracy for Subdural Hemorrhage (SDH); 34.89%, 18.94%, 28.37% and 33.67% higher accuracy for Subarachnoid Hemorrhage (SAH); 29.45%, 34.67%, 45.78% and 19.56% higher accuracy for Intracerebral Hemorrhage (ICH); 18.49%, 19.83%, 22.89% and 12.05% higher accuracy for Intraventricular Hemorrhage (IVH) compared with existing ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet methods respectively.

Table 1

Performance of accuracy analysis

	Methods
Types of acute ICH	ICH-DC-	ICH-DC-CNN-	ICH-DC-S-	ICH-DC-CNN-	ICH-DC-ResNet-DenseNet-
	CNN	ResNet-50	3D-CNN	AlexNet	XGBoost (Proposed)
Epidural Hemorrhage (EDH)	72	75	62	82	99
Subdural Hemorrhage (SDH)	80	72	79	62	98
Subarachnoid Hemorrhage (SAH)	74	88	67	75	97.32
Intracerebral Hemorrhage (ICH)	64	76	73	67	98
Intraventricular Hemorrhage (IVH)	74	56	80	79	96

Table 2 depicts the precision analysis, the performance of the proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 32.15%, 18.29%, 23.65%, and 14.20% higher precision for Epidural Hemorrhage (EDH); 55.25%, 42.57%, 12.58% and 22.18% higher precision for Subdural Hemorrhage (SDH); 25.01%, 26.58%, 21.48% and 41.28% higher precision for Subarachnoid Hemorrhage (SAH); 30.28%, 20.17%, 23.58% and 16.58% higher precision for Intracerebral Hemorrhage (ICH); 21.48%, 47.25%, 36.25% and 20.14% higher precision for Intraventricular Hemorrhage (IVH) estimated with existing ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet methods respectively.

Table 2

Performance of precision analysis

	Methods
Types of acute ICH	ICH-DC-	ICH-DC-CNN-	ICH-DC-S-	ICH-DC-CNN-	ICH-DC-ResNet-DenseNet-
	CNN	ResNet-50	3D-CNN	AlexNet	XGBoost (Proposed)
Epidural Hemorrhage (EDH)	60	75	65	82	98.52
Subdural Hemorrhage (SDH)	55	82	75	75	97.28
Subarachnoid Hemorrhage (SAH)	63	64	73	86	96.48
Intracerebral Hemorrhage (ICH)	72	72	81	59	98.17
Intraventricular Hemorrhage (IVH)	86	84	64	62	99.36

Table 3 depicts the f-measure analysis, the performance of the proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 40.25%, 52.01%, 36.28%, and 28.52% higher f-measure for Epidural Hemorrhage (EDH); 33.21%, 24.87%, 33.25% and 24.56% higher f-measure for Subdural Hemorrhage (SDH); 23.65%, 19.36%, 20.31% and 39.85% higher f-measure for Subarachnoid Hemorrhage (SAH); 37.52%, 34.28%, 32.17% and 35.28% higher f-measure for Intracerebral Hemorrhage (ICH); 25.15%, 17.08%, 25.18% and 33.28% higher f-measure for Intraventricular Hemorrhage (IVH) estimated with existing ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet methods respectively.

Table 3

Performance of F-measure analysis

	Methods
Types of acute ICH	ICH-DC-	ICH-DC-CNN-	ICH-DC-S-	ICH-DC-CNN-	ICH-DC-ResNet-DenseNet-
	CNN	ResNet-50	3D-CNN	AlexNet	XGBoost (Proposed)
Epidural Hemorrhage (EDH)	55	85	73	64	99
Subdural Hemorrhage (SDH)	85	78	65	72	96
Subarachnoid Hemorrhage (SAH)	62	86	52	74	97
Intracerebral Hemorrhage (ICH)	75	63	85	52	98.15
Intraventricular Hemorrhage (IVH)	89	55	72	61	97.25

Table 4 depicts the recall analysis, the performance of the proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 25.36%, 63.25%, 52.14%, and 30.28% higher recall for Epidural Hemorrhage (EDH); 53.28%, 23.58%, 34.52% and 30.75% higher recall for Subdural Hemorrhage (SDH); 52.17%, 56.28%, 45.28% and 37.59% higher recall for Subarachnoid Hemorrhage (SAH); 37.85%, 41.28%, 50.28% and 42.85% higher recall for Intracerebral Hemorrhage (ICH); 31.25%, 33.28%, 22.18% and 39.74% higher recall for Intraventricular Hemorrhage (IVH) estimated with existing ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet methods respectively.

Table 4

Performance of recall analysis

	Methods
Types of acute ICH	ICH-DC-	ICH-DC-CNN-	ICH-DC-S-	ICH-DC-CNN-	ICH-DC-ResNet-DenseNet-
	CNN	ResNet-50	3D-CNN	AlexNet	XGBoost (Proposed)
Epidural Hemorrhage (EDH)	59	62	75	82	98.25
Subdural Hemorrhage (SDH)	62	59	86	75	99.85
Subarachnoid Hemorrhage (SAH)	85	75	55	63	96.32
Intracerebral Hemorrhage (ICH)	64	86	72	52	99.17
Intraventricular Hemorrhage (IVH)	73	68	63	87	97.86

Table 5 depicts the specificity analysis, the performance of the proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 52.38%, 45.28%, 29.67%, and 22.85% higher specificity for Epidural Hemorrhage (EDH); 32.65%, 44.28%, 32.65% and 23.86% higher specificity for Subdural Hemorrhage (SDH); 34.28%, 19.67%, 16.78% and 19.78% higher specificity for Subarachnoid Hemorrhage (SAH); 25.39%, 45.28%, 36.28% and 19.66% higher specificity for Intracerebral Hemorrhage (ICH); 22.18%, 23.85%, 20.75% and 35.28% higher specificity for Intraventricular Hemorrhage (IVH) estimated with existing ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet methods respectively.

Table 5

Performance of specificity analysis

	Methods
Types of acute ICH	ICH-DC-	ICH-DC-CNN-	ICH-DC-S-	ICH-DC-CNN-	ICH-DC-ResNet-DenseNet-
	CNN	ResNet-50	3D-CNN	AlexNet	XGBoost (Proposed)
Epidural Hemorrhage (EDH)	52	63	85	74	99
Subdural Hemorrhage (SDH)	86	75	55	69	98
Subarachnoid Hemorrhage (SAH)	85	63	75	81	97
Intracerebral Hemorrhage (ICH)	75	63	72	85	96
Intraventricular Hemorrhage (IVH)	63	56	62	56	99.36

Table 6 depicts the sensitivity analysis, the performance of the proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 24.15%, 25.18%, 36.75%, and 39.48% higher sensitivity for Epidural Hemorrhage (EDH); 25.96%, 23.84%, 19.74% and 25.84% higher sensitivity for Subdural Hemorrhage (SDH); 21.08%, 29.45%, 32.18% and 24.58% higher sensitivity for Subarachnoid Hemorrhage (SAH); 15.28%, 20.75%, 32.58% and 36.28% higher sensitivity for Intracerebral Hemorrhage (ICH); 29.45%, 27.36%, 26.19% and 34.58% higher sensitivity for Intraventricular Hemorrhage (IVH) estimated with existing ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet methods respectively.

Table 6

Performance of sensitivity analysis

	Methods
Types of acute ICH	ICH-DC-	ICH-DC-CNN-	ICH-DC-S-	ICH-DC-CNN-	ICH-DC-ResNet-DenseNet-
	CNN	ResNet-50	3D-CNN	AlexNet	XGBoost (Proposed)
Epidural Hemorrhage (EDH)	63	85	76	72	96
Subdural Hemorrhage (SDH)	55	74	89	62	99
Subarachnoid Hemorrhage (SAH)	80	63	62	85	98
Intracerebral Hemorrhage (ICH)	72	69	72	62	97
Intraventricular Hemorrhage (IVH)	85	59	61	73	96.28

Table 7 depicts the error rate, the performance of the proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 55.26%, 45.28%, 39.45%, and 42.38% lower error rate for Epidural Hemorrhage (EDH); 35.45%, 42.65%, 41.85% and 21.05% lower error rate for Subdural Hemorrhage (SDH); 39.45%, 42.15%, 45.16% and 52.15% lower error rate for Subarachnoid Hemorrhage (SAH); 23.15%, 45.15%, 32.61% and 37.19% lower error rate for Intracerebral Hemorrhage (ICH); 23.18%, 50.19%, 43.28% and 47.52% lower error rate for Intraventricular Hemorrhage (IVH) estimated with existing ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet methods respectively.

Table 7

Performance of error rate analysis

	Methods
Types of acute ICH	ICH-DC-	ICH-DC-CNN-	ICH-DC-S-	ICH-DC-CNN-	ICH-DC-ResNet-DenseNet-
	CNN	ResNet-50	3D-CNN	AlexNet	XGBoost (Proposed)
Epidural Hemorrhage (EDH)	28	25	38	18	1
Subdural Hemorrhage (SDH)	20	28	21	38	2
Subarachnoid Hemorrhage (SAH)	26	12	33	25	2.68
Intracerebral Hemorrhage (ICH)	36	24	27	33	2
Intraventricular Hemorrhage (IVH)	26	44	20	21	4

Table 8 depicts the computational time for Intracranial Haemorrhage (ICH) subtype detection and classification. Then, the computational time of proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 2.89%, 6.59%, 5.43% and 5.06% lower computational time than the existing ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet methods respectively.

Table 8

Computational time

Methods
ICH-DC-	ICH-DC-CNN-	ICH-DC-S-	ICH-DC-CNN-	ICH-DC-ResNet-DenseNet-
CNN	ResNet-50	3D-CNN	AlexNet	XGBoost (Proposed)
0.52	0.48	0.19	0.35	0.5

ROC Curve of ICH-DC-ResNet-DenseNet-XGBoost model is represented in Fig. 4. Here, the proposed ICH-DC-ResNet-DenseNet-XGBoost method provides 3.157%, 4.158%, 35.94%, 35.56, 27.95% and 2.098% higher AUC than the existing methods like ICH-DC-CNN, ICH-DC-CNN-ResNet-50, ICH-DC-S-3D-CNN, and ICH-DC-CNN-AlexNet respectively.

Fig. 4

ROC curve performance.

5 Conclusions

Intracranial hemorrhages are considered as an acute brain disease that may lead to death if not diagnosed in its early stage. It is essential to detect the type of hemorrhage for early treatment and diagnosis. In this paper, an automatic detection and classification of acute ICH is carried out by using an efficient feature extraction and classification framework. At first, pre-processing is performed to remove the outliers from the raw CT image taken from publicly available Physionet repository dataset. The relevant features essential to classify the subtypes of acute ICH are extracted by adopting a combined ResNet-DenseNet architecture. Finally, the acute ICH is categorized into 5 subtypes using XGBoost classifier based on the critical features extracted from the CT image. Experimental outcomes display that the ResNet-DenseNet-XGBoost framework attains better performance using high classification accuracy and less misclassified error when compared to existing state of art intracranial hemorrhage detection and classification methods. The formulation increment to non-convex optimization problems, for eg, deep neural network is an attractive future prospect. In future, we aim to solve the convergence scheme for the class of non-convex loss functions.

Data availability statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

Funding Information

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

References

Kim

K.H.

, Koo

H.W.

, Lee

B.J.

, Yoon

S.W.

, Sohn

M.J.

Cerebral hemorrhage detection and localization with medical imaging for cerebrovascular disease diagnosis and treatment using explainable deep learning, Journal of the Korean Physical Society (2021), 1–7.

Rava

R.A.

, Seymour

S.E.

, LaQue

M.E.

, Peterson

B.A.

, Snyder

K.V.

, Mokin

, Waqas

, Hoi

, Davies

J.M.

, Levy

E.I.

and Siddiqui

A.H.

, Assessment of an Artificial Intelligence Algorithm for Detection of Intracranial Hemorrhage, World Neurosurgery 150 (2021), 209–e217.

Medani

, Agrawal

Neuroprotection in paediatric traumatic brain injury, Paediatrics and Child Health (2021).

Raghavendra

, Gudigar

, Vidhya

, Rao

B.N.

, Sabut

, Wei

J.K.E.

, Ciaccio

E.J.

and Acharya

U.R.

, Novel and accurate non-linear index for the automated detection of haemorrhagic brain stroke using CT images, Complex & Intelligent Systems 7(2) (2021), 929–940.

Berikol

G.B.

, Berikol

, Bozdereli

Artificial intelligence in neuro, head, and neck surgery. In Artificial Intelligence in Precision Health, Academic Press (2020), 393–404.

Alves

, Wancket

, Metz

Current considerations in medical device pathology. In Biocompatibility and Performance of Medical Devices, Woodhead Publishing (2020), 491–543.

Johnson

E.G.

, Prabhakar

, Mooney

L.N.

and Ghetti

, Neuroimaging the sleeping brain: Insight on memory functioning in infants and toddlers, Infant Behavior and Development 58 (2020), 101427.

Kaliuzhna

, Kirschner

, Carruzzo

, Hartmann-Riemer

M.N.

, Bischof

, Seifritz

, Tobler

P.N.

and Kaiser

, Clinical, behavioural and neural validation of the PANSS amotivation factor, Schizophrenia Research 223 (2020), 8–45.

Watanabe

, Tanaka

, Nishida

, Takahashi

, Fujiwara

and Fujiwara

, A.Arisawa, H. Yano, N. Tomiyama, H. Nakamura and K. Todo, Improvement of the diagnostic accuracy for intracranial haemorrhage using deep learning–based computer-assisted detection, Neuroradiology 63(5) (2021), 713–720.

10.

Reddy

and Prakash

, Enhanced key establishment technique for secure data access in cloud, In 2019 International Conference on Issues and Challenges in Intelligent Computing Techniques (ICICT)) 1 (2019), pp. 1–4. IEEE.

11.

Mangesius

, Janjic

, Steiger

, Haider

, Rehwald

, Knoflach

, Widmann

, Gizewski

and Grams

, Dual-energy computed tomography in acute ischemic stroke: state-of-the-art, European Radiology 31(6) (2021), 4138–4147.

12.

Gaddam

D.S.

, Dattwyler

, Fleiter

T.R.

and Bodanapally

U.K.

, Principles and Applications of Dual Energy Computed Tomography in Neuroradiology, In, Seminars in Ultrasound, CT and MRI WB Saunders 42(5) (2011), 418–433.

13.

Zhang

, Cao

, Kong

, Mao

, Huang

, He

, Pan

, Li

and Lu

, Clinical and hematological characteristics of 88 patients along COVID-19, International Journal of Laboratory Hematology 42(6) (2020), 780–787.

14.

Mangaonkar

A.A.

and Patnaik

M.M.

, Hereditary predisposition to hematopoietic neoplasms: when bloodline matters for blood cancers. In, Mayo Clinic Proceedings Elsevier 95(7) (2020), 1482–1498.

15.

https://ieeexplore.ieee.org/document/9670691

16.

https://ieeexplore.ieee.org/document/8437157

17.

Francisco

M.Z.

, Altmayer

, Verma

, Watte

, de Campos Brandao

, Barros

M.C.

, Mohammed

T.L.

and Hochhegger

, Appropriateness of Computed Tomography with Ultrasound for Abdominal Complaints in the Emergency Department, Current Problems in Diagnostic Radiology 50(6) (2021), 799–802.

18.

Wang

, Shen

, Yang

, Lan

, Xu

, Wang

, Zhang

and Han

, A deep learning algorithm for automatic detection and classification of acute intracranial hemorrhages in head CT scans, NeuroImage: Clinical 32 (2021), 102785.

19.

Alfaer

N.M.

, Aljohani

H.M.

, Abdel-Khalek

, Alghamdi

A.S.

, Mansour

R.F.

Fusion-Based Deep Learning with Nature-Inspired Algorithm for IntracerebralHaemorrhage Diagnosis, Journal of Healthcare Engineering 2022.

20.

Alshabibi

A.S.

, Mo’ayyad

E.S.

, Tapia

K.A.

, Heard

and Brennan

P.C.

, Impact of Hours Awake and Hours Slept at Night on Radiologists’ Mammogram Interpretations, Journal of the American College of Radiology 18(5) (2021), 730–738.

21.

McCormick

E.M.

, Lott

M.T.

, Dulik

M.C.

, Shen

, Attimonelli

, Vitale

, Karaa

, Bai

, Pineda-Alvarez

D.E.

, Singh

L.N.

and Stanley

C.M.

, Specifications of the ACMG/AMP standards and guidelines for mitochondrial DNA variant interpretation, Human Mutation 41(12) (2020), 2028–2057.

22.

Rava

R.A.

, Seymour

S.E.

, LaQue

M.E.

, Peterson

B.A.

, Snyder

K.V.

, Mokin

, Waqas

, Hoi

, Davies

J.M.

, Levy

E.I.

and Siddiqui

A.H.

, Assessment of an Artificial Intelligence Algorithm for Detection of Intracranial Hemorrhage, World Neurosurgery 150 (2021), 209–e217.

23.

Wang

, Shen

, Yang

, Lan

, Xu

, Wang

, Zhang

and Han

, A deep learning algorithm for automatic detection and classification of acute intracranial hemorrhages in head CT scans, Neuro Image: Clinical 32 (2021), 102785.

24.

Battineni

Machine Learning and Deep Learning Algorithms in the Diagnosis of Chronic Diseases, In Machine Learning Approaches for Urban Computing Springer, Singapore (2021), 141–164.

25.

Shajin

F.H.

, Rajesh

and Raja

M.R.

, An efficient VLSI architecture for fast motion estimation exploiting zero motion prejudgment technique and a new quadrant-based search algorithm in HEVC, Circuits, Systems, and Signal Processing 41(3) (2022), 1751–1774.

26.

Rajesh

, Shajin

F.H.

An efficient hybrid tunicate swarm algorithm and radial basis function searching technique for maximum power point tracking in wind energy conversion system, Journal of Engineering, Design and Technology (2021).

27.

Shajin

F.H.

and Rajesh

, FPGA realization of a reversible data hiding scheme for 5G MIMO-OFDM system by chaotic key generation-based paillier cryptography along with LDPC and its side channel estimation using machine learning technique, Journal of Circuits, Systems and Computers 31(05) (2022), 2250093.

28.

Rajesh

, Muthubalaji

, Srinivasan

and Shajin

F.H.

, Leveraging a Dynamic Differential Annealed Optimization and Recalling Enhanced Recurrent Neural Network for Maximum Power Point Tracking in Wind Energy Conversion System, Technology and Economics of Smart Grids and Sustainable Energy 7(1) (2022), 1–5.

29.

Badža

M.M.

and Barjaktarović

M.Č.

, Classification of brain tumors from MRI images using a convolutional neural network, Applied Sciences 10(6) (2020), 1999.

30.

Yadav

S.S.

and Jadhav

S.M.

, Deep convolutional neural network based medical image classification for disease diagnosis, Journal of Big Data 6(1) (2019), 1–18.

31.

Ker

, Bai

, Lee

H.Y.

, Rao

and Wang

, Automated brain histology classification using machine learning, Journal of Clinical Neuroscience 66 (2019), 239–245.

32.

Anupama

C.S.S.

, Sivaram

, Lydia

E.L.

, Gupta

, Shankar

Synergic deep learning model–based automated detection and classification of brain intracranial hemorrhage images in wearable networks, Personal and Ubiquitous Computing (2020), 1–10.

33.

Zhang

, Chen

, Xu

, Cho

P.C.

, Nan

, Zhou

, Lv

, Li

and Xie

, Lesion synthesis to improve intracranial hemorrhage detection and classification for CT images, Computerized Medical Imaging and Graphics 90 (2021), 101929.

34.

Karki

, Cho

, Lee

, Hahm

M.H.

, Yoon

S.Y.

, Kim

, Ahn

J.Y.

, Son

, Park

S.H.

, Kim

K.H.

and Park

, CT window trainable neural network for improving intracranial hemorrhage detection by combining multiple settings, Artificial Intelligence in Medicine 106 (2020), 101850.

35.

Mansour

R.F.

, Escorcia-Gutierrez

, Gamarra

, Díaz

V.G.

, Gupta

, Kumar

Artificial intelligence with big data analytics-based brain intracranial hemorrhage e-diagnosis using CT images, Neural Computing and Applications (2021), 1–13.

36.

Hausman-Kedem

, Malinger

, Modai

, Kushner

S.A.

, Shiran

S.I.

, Ben-Sira

, Roth

, Constantini

, Fattal-Valevski

and Ben-Shachar

, Monogenic Causes of Apparently Idiopathic Perinatal Intracranial Hemorrhage, Annals of Neurology 89(4) (2021), 813–822.

37.

, Gao

, Yin

, Guo

, Zhao

, Lu

, Wang

, Bai

, Cao

, Song

and Zhang

, Precise diagnosis of intracranial hemorrhage and subtypes using a three-dimensional joint convolutional and recurrent neural network, European Radiology 29(11) (2019), 6191–6201.

38.

Patel

, Van De Leemput

S.C.

, Prokop

, Van Ginneken

and Manniesing

, Image level training and prediction: intracranial hemorrhage identification in 3D non-contrast CT, IEEE Access 7 (2019), 92355–92364.

39.

, Chen

, He

, Zhang

, Chen

, Li

and Gao

, Automatic segmentation of intracerebral hemorrhage in CT images using encoder–decoder convolutional neural network, Information Processing & Management 57(6) (2020), 102352.

40.

Lee

J.Y.

, Kim

J.S.

, Kim

T.Y.

and Kim

Y.S.

, Detection and classification of intracranial haemorrhage on CT images using a novel deep-learning algorithm, Scientific Reports 10(1) (2020), 1–7.

41.

Sage

and Badura

, Intracranial hemorrhage detection in head CT using double-branch convolutional neural network, support vector machine, and random forest, Applied Sciences 10(21) (2020), 7577.

42.

Singh

S.P.

, Wang

, Gupta

, Gulyas

and Padmanabhan

, Shallow 3D CNN for detecting acute brain hemorrhage from medical imaging sensors, IEEE Sensors Journal 21(13) (2020), 14290–14299.

43.

Dhillon

and Verma

G.K.

, Convolutional neural network: a review of models, methodologies and applications to object detection, Progress in Artificial Intelligence 9(2) (2020), 85–112.

44.

Devan

, Khare

An efficient XGBoost–DNN-based classification model for network intrusion detection system, Neural Computing and Applications (2020), 1–6.

45.

Choi

, Shin

N.Y.

, Jang

, Ahn

K.J.

and Kim

B.S.

, Dual-energy CT for differentiating acute intracranial hemorrhage from contrast staining or calcification: a meta-analysis, Neuroradiology 62(12) (2020), 1617–1626.

46.

Yu Koltun

F.V.

, Funkhouser

Dilated residual networks in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017), 472–480.