Artificial intelligence with big data analytics-based brain intracranial hemorrhage e-diagnosis using CT images

Due to the fast development of medical imaging technologies, medical image analysis has entered the period of big data for proper disease diagnosis. At the same time, intracerebral hemorrhage (ICH) becomes a serious disease which affects the injury of blood vessels in the brain regions. This paper presents an artificial intelligence and big data analytics-based ICH e-diagnosis (AIBDA-ICH) model using CT images. The presented model utilizes IoMT devices for data acquisition process. The presented AIBDA-ICH model involves graph cut-based segmentation model for identifying the affected regions in the CT images. To manage big data, Hadoop Ecosystem and its elements are mainly used. In addition, capsule network (CapsNet) model is applied as a feature extractor to derive a useful set of feature vectors. Finally, the presented AIBDA-ICH model makes use of the fuzzy deep neural network (FDNN) model to carry out classification process. For validating the superior performance of the AIBDA-ICH method, an extensive set of simulations were performed and the outcomes are examined under diverse aspects. The experimental values pointed out the improved e-diagnostic performance of the AIBDA-ICH model over the other compared methods with the precision and accuracy of 94.96% and 98.59%, respectively.


Introduction
Globally, several countries experience a drastic rise in the huge number of patients. This increase in patient cases is limiting the person to access the caretakers or physician for better treatment. With the latest technological developments, enhancing the performance of biomedicine and healthcare methods is one of the difficult processes. As pointed out by [1], the presently obtainable process for medical treatment, observing, supervision, and administration is performed manually by medical staff. Technically, giving better treatment to persons with inexpensive, over limited time, and addressing the insufficiency of nurses are the major problems. In previous times, the growth of internet of things (IoT) and wearables has enhanced treatment by monitoring the patients from remote areas [2]. Currently, IoT gadgets act as an important part of healthcare application areas particularly in the detection and diagnosis of many diseases in smart city and navigating implementation method, effective traffic prioritizing system in IoT gadgets, strengthen the Internet asset content, and handle a lot of Internet connection by information centric network (ICN). Figure 1 shows the pipeline of big data [3]. Brain hemorrhage is a kind of stroke which is generally instigated by an artery in the brain such as stroke burst and bleeding in the nearby tissues. The incessant bloodstream from the concerned tissue destroys brain cells. When the disease is not precisely diagnosed and delayed the treatment process, it might lead to mortality or disabilities forever. The major reason for brain hemorrhage is adequate alcohol consumption, smoking, high blood pressure, etc., whereas inheritance is further assumed as the main reason for brain hemorrhage. The computed tomography (CT) scanning is investigated by the radiotherapists to forecast intracerebral hemorrhage (ICH) and detect injured areas. ICH is relevant to the occurrence of bleeding inside the intracranial vault. The probable causes comprise venous infarction, tumor, vascular abnormality, trauma effect, cerebral aneurysm, and therapeutic anticoagulation. However, the effective causes of hemorrhage pose the main risk. Consequently, a precise and quick diagnosis is critical for processing treatment, and it attains successful outcomes. The ICH diagnoses depend upon patient's past medical details, examination of physical and non-contrast CT of the brain region. The CT investigation allows bleeding location and denotes the main cause of ICH.
Several issues have existed in the detection and classification of ICH such as time-consuming, lengthier process, need highly professional radiotherapists, etc. Therefore, there is an important requirement for a computer-aided diagnosis method to support the professionals. Nonetheless, the accuracy of automatic hemorrhage recognition must be considered high for medicinal reasons. The automated prioritization of imaging researches by the use of computerized techniques has the possibility of detecting ICH at the beginning stage, eventually resulting in enhanced medical results. Such a quality enhancement tool can be utilized to automatically managing the preference for interpreting the imaging study with assumed ICH and assist optimized radiology plan. The ML and computer vision are the set of methods used to teach computers for learning and detecting patterns.
Recently, deep learning (DL) models have gained significant attention for image segmentation and classification. CNN has become more popular because of the efficiency and reliability, which becomes major factor in medicinal diagnosis assistance. Generally, CT scan is in a three-dimensional structure which consists of a stack of two-dimensional slices. Therefore, the functioning of image voxel is probable; however, it might need high computation complexity. A method to prevent the latter is to compute slices separately or apply three-dimensional context in an easier manner. Numerous DL techniques for ICH diagnosis have been presented in recent times [4,5]. Distinct researches resolve 1 class recognition regarding individual classes of ICH existing in CT scan, multi-class classifier differentiating ICH subtypes, or ICH pixel part recognition inside the single image. The familiarity of pre-trained convolutional neural network (CNN) methods is notable in Fig. 1 Pipeline of big data [3] several models such as ResNet-18, Alex Net, VGG, and Mobile Net. This paper presents an artificial intelligence and big data analytics-based ICH e-diagnosis (AIBDA-ICH) model using CT images. The presented model exploits IoMT devices for data acquisition process. The presented AIBDA-ICH model involves graph cut-based segmentation model for identifying the affected regions in the CT images. For managing big data, Hadoop Ecosystem and its corresponding components are extremely utilized. Moreover, Capsule Network (CapsNet) model is applied as a feature extractor to derive a useful set of feature vectors. Finally, the presented AIBDA-ICH model uses fuzzy deep neural network (FDNN) model to perform classification process. To assess the betterment of the AIBDA-ICH model, a series of experimentation was performed and the outcomes are inspected under varied aspects.

Literature review
Prevedello et al. [6] proposed two methods depending upon CNN. Among these two methods, the initial one focuses on ICH prediction, hydrocephalus, and mass effects during ICH scanning, while the subsequent one is employed for the detection of malicious acute infarctions. Later, entire CT scans are employed for training, testing, and verification. They are relevant to CNN with recurrent neural network (RNN) that is proposed for ICH detection. Grewal et al. [7] proposed a 40-layer CNN, called Dense Net with bi-directional long short-term memory (Bi-LSTM) layer for predicting ICH. Furthermore, three supplementary functions are proposed for the dense convolutional block to perform the binary segmentation for ICH regions. These processes consist of single convolutional and deconvolution layer to compute the sample feature mapping to original image size. Ye et al. [8] proposed a three-dimensional joint CNN-RNN to predict and classify ICH region. VGG-16 is applied as CNN technique, and bi-directional gated recurrent unit (GRU) is employed as RNN method. The RNN layer performs identical to slice interpolation method given by Lee et al. [9]; though, it exhibits flexibility using closest slice in the classifier.
Jnawalia et al. [10] proposed an ensemble of three varied CNN systems for ICH prediction. The CNN techniques are based on the structures of AlexNet and Goo-gleNet that are extended to three-dimension method through the consumption of whole slices. Additionally, the smallest number of variables and filter particulars are assumed. In Arbabshirani et al. [11], an ensemble of threedimension CNN methods with the input is implemented and calculated by utilization of retrospective and prospective CT scans. The novel CNN technique to identify subarachnoid hemorrhages, intraparenchymal, and epidural or subdural over non-contrast CT is given in [12]. Also, Majumdar et al. [13] proposed a DCNN for performing simultaneous feature extraction and classification, discarding several hand-crafted processes. Although numerous ICH diagnosis methods are obtainable in the studies, a minimum number of researches have concentrated on DL techniques. Additionally, the ICH diagnoses methods in the wearable network platform become vital to supporting medicinal facilities in remote areas.
Distinct methods are presented for telehealth purposes depending upon IoT gadgets. For instance, Abdelaziz [14] introduced an ML technique depending upon linear regression and NN for chronic kidney disease (CKD) diagnoses and forecasting. Al-Majeed et al. [15] projected a home telehealth method depending upon IoT gadgets. Dwivedi et al. [16] presented blockchain technology for information security in IoT gadgets. Firouzi [17] recommended intelligent sensors for healthcare applications. Hassanalieragh [18] resolved the present problems to face IoT gadgets and proposed results to manage this critical problem. Jabbar [19] introduced an IoT semantic based interoperability method (IoT-SIM) to establish semantic interoperability between heterogeneous IoT gadgets in the healthcare field. Maktoubian and Ansari [20] presented integrity observing method to ensure privacy and security of healthcare IoT gadgets. Mutlag et al. [21] proposed an efficient study for fog computing in healthcare IoT gadgets. Shakeel [22] presented a deep Q-network for constraining malicious attacks in healthcare IoT gadgets.

System architecture
The overall system architecture of the proposed AIBDA-ICH model is depicted in Fig. 2. The e-diagnosis model initially utilizes a set of IoMT devices to acquire patient's data. The IoMT devices transfer the data to the cloud server for remote diagnosis. In addition, to handle big data, Hadoop ecosystem is utilized. In order to manage Big Data, Hadoop Ecosystem and its components are extremely utilized. On the distributed background, Hadoop is the type of open-source structure which allows stakeholders for storing and processing the Big Data on computer clusters by utilizing simple programming methods. For thousands of nodes in single server, it can be exhibited to contain improved scalability and fault tolerance. The three important elements of Hadoop are MapReduce, Hadoop Distributed File System (HDFS), and Hadoop YARN. Based on Google File System (GFS), HDFS is demonstrated. It can be shown as structure of variety such as master or slave somewhere the master has additional data nodes that are known as actual data and different name node that is known as metadata.

Hadoop map reduce
In order to offer the massive scalability on thousand Hadoop clusters, the Hadoop Map Reduce is utilized that is called programming structure at Apache Hadoop heart. For processing enormous information over enormous clusters, MapReduce is utilized. The two important phases are controlled from MapReduce job models like Reduce and Map phase. All the phases include pairs such as key-value which is input as well as output; definitely from the file scheme, both outputs and input of job are protected. This framework satisfies classification, adjusting it, and fails to the task re-implementation. The framework of MapReduce involves one slave node manager and single master resource manager to every cluster node.

Hadoop YARN
The Hadoop YARN is a method utilized to manage clusters. In the improved experience of initial Hadoop generation, it can be shown as second Hadoop generation that performs as an important feature [23]. Over Hadoop clusters, to deals with security, reliable functions, and data governance devices, YARN performs as the central structure and resource manager. To control the big data, other platform tools and elements are installations over Hadoop framework. The presented ICH e-diagnosis method contains a series of processes such as preprocessing, graph-cut-based segmentation, CapNet-based feature extraction, and FDNN-based classification. Firstly, contrast enhancement processes take place to improve the quality of the image. It is followed by graph cut-based segmentation model which is implemented for identifying the affected brain regions in the CT scan. Next to that, CapsNet model is employed to determine an appropriate set of feature vectors from the segmented image. At last, FDNN model is applied to assign proper class labels to the input test images. Thus, the utilization of AI and DL models helps to effectively diagnose the ICH of remote patients. The processes involved in these techniques are discussed briefly in the succeeding sections.

Graph-cut-based Segmentation
The graph cuts-based segmentation technique is implemented to detect the affected brain regions, and the key is to design a graph G ¼ V; E; W ð Þ , where V; E; W are the group of vertices, edges, and weights on the edges. According to graph G, the max-flow/min-cut technique is utilized for finding the minimal capacity over every cut of graph G. The capacity of cut is provided as: where w a; b ð Þ is utilized as a weighting of the edge connected neighboring pixels a and b and for building the graph G. To seed pixels connected to ends S and T, the weightings must be fixed to large values ! to make sure that the edge not to be cut [24]. The weighting w a; b ð Þ for edge e ab connected neighboring pixe1 a and b can be determined as: where w c a; b ð Þ is a restriction term and w bs a; b ð Þ refers the edge-based term. The local region-based w r 1 a; b ð Þ is determined depending upon the local intensity data approximatively the classifier outcome. The remote-local region based on w r 2 a; b ð Þ is determined by choosing neighboring points in trained information regions depending upon Euclidean distance.
Especially, the four terms in Eq.
(2) are determined as: w a a; b where parameter r in Eqs. (3)-(6) is utilized to adjust the similarity measurement among pixels a and b. In addition, D a; b ð Þ is determined depending upon the distanced transform of classification outcome as given below.
where Dist a ð Þ is the minimum distance of pixel p to the restricted contour area.  5) and (6), the terms K r 1 a; b ð Þ and K r 2 a; b ð Þ take the similar method K r a; b ð Þ determined in Eq. (8), however with several f l a r a ð Þ and f l b r b ð Þ for Eqs. (5) and (6).
where the subscript r is the abbreviation of r À 1 and r À 2: In Eq. ð Þ is the local region-based term, and w r 2 a; b ð Þ is the remote-local region-based term. The radius to compute f l a r 1 a ð Þ and f l b r 1 b ð Þ is fixed as R 1 ¼ R=3 with R be the width of brush utilized for drawing the narrowband and the number n 1 is utilized to compute the remote-local version as fixed by n 1 ¼ 2 Ã 2 Ã R 1 þ 1 ð Þ for capturing the larger region.

CapsNet based feature extraction
Once the images are segmented, CapsNet model is employed for feature extraction. Hinton et al. altered the viewpoint of image investigation from invariance to equivariance with the assumption of invariance [25]. These are the two main elements of visual depiction. Invariance is implemented for a particular process whereas equivariance is developed for several processes like orientation, pose, or location. Alternatively, equivariance remains almost every data required for image representation. The CapsNet method is a recent topic involved in the field of computer vision. The fundamental unit of CapsNet is a capsule that comprises set of structured neurons. The extent of capsule is based on invariance, whereas several features to rebuild the image are the measure of equivariance. The capsules generate vectors of similar magnitude still by distinct orientations. The orientation of vector denotes variables, especially the data of the property maintained from the image. Figure 3 demonstrates the structure of capsule network [26].
A regular NN requires more layers for raising accuracy and detail, by CapsNet, an individual layer can nestle additional layer. The capsule efficiently denotes distinct kinds of visual data, is termed as instantiation variables, like pose, that is integration of size, orientation, and position. The outcome of the capsule is a represent vector that is sending to a layer overhead to compare its proper parental. The result of capsules i is assumed to u i , and conversion matrix W ij is employed to estimate the capsule outcome for the predictive parental capsule j via converting u i to predictive vectorÛjji: whereÛjji denotes predictive vector of output jth capsules in high level calculated via capsules i in below layer, and W ij represents weight matrix which is required to learn in backward pass. The variable s j denotes the weighted amount entire predictive vector u jji in which c ij denotes coupling coefficient computed via dynamic routing procedure to define degree of configuration among the capsule in below layer and parental capsule. Dynamic routing is the procedure of generating parental capsule via coupling the capsules by ''routing by agreement'' method. This relation is not designed via ''max pooling'' of typical CNN. Different from max pooling, whole required data from detail are maintained, therefore achievement increased via overlap image. The dimension of the capsule rises when hierarchy increases.
The activation function is known as squashing shrink the last resultant vector to 0 when they are tiny and unit vector when they are high and generate capsules length. The activation vector v j is determined by subsequent nonlinear squashing function.
The c ij is calculated as the softmax of b ij . The coupling coefficient is determined as the degree of configuration among the capsules and parental capsules.
b ij is the resemblance score which considers the account both likeliness and features properties rather than just likeliness in neurons.

FDNN-based classification
In this stage, the generated feature vectors from the CapsNet are fed into the FDNN model for classification process. The FDNN consists of 4 learning portions. In a nutshell, the input data follow two ways to represent fuzzy logic (FL) and neurons. So, the illustrations from these two aspects are integrated with the fusion part (i.e., green part). Similarly, the fused data of the initial layer are more serially converted by creating red layer at last. The red layer is the process determined part related to the classification of allocating data points to distinct classes [27]. Let l represent layer number, a l ð Þ i denote input of the ith node and o l ð Þ i indicate corresponding output. In every part, l represents presently discussed layer. Part I-FL depiction: All nodes in the input layer are linked to many membership functions which allocate linguist labels to every input parameter. Now, the input parameter represents the 1D of input vector. It estimates the degree that an input node belonging to specific fuzzy set. In these layers, ith fuzzy neuron u i Á ð Þ: R ! 0; 1 ½ mapping kth input as fuzzy degree is given by The Gaussian membership function by mean l and variance r 2 are used in this technique. The fuzzy rule layer achieves 'AND' FL operations, such as o 8j 2 X i ; in which X i denotes class of node on the l À 1 ð Þth layer which is connected to i. Then, the l À 1 ð Þth layer represents input layer. The resultant part is fuzzy degree.
Part II-Neural depiction: This part utilizes neural learning method to convert the input into few higher level depictions. This layer is fully connected (FC) which implies all nodes in l ð Þth layer is linked to each node in and is given by Let w l ð Þ i and b l ð Þ i denote the weights and bias linking to node i over lth layer.
Part III-Fusion part: The fused method used is stimulated through multi-modal learnings. It is believed that feature extraction from individual views is not appropriate for capturing difficult structures of high content information. Therefore, this technique adequately generates multiple features from several characteristics and combines them to higher level depiction for classifier. The fuzzy and neural parts are utilized in FDNN to search optimum depictions via decreasing noises of input data. To understand better this technique, reader assumed the output of fuzzy part as feature instead of actual fuzzy fundamentals. Furthermore, neural and fuzzy learning part is designed in NN. Consequently, it is very spontaneous to design the fusion feature phase in NN set. Here, the broadly utilized multi-model NN structures [28] are preferred to integrate neural and fuzzy depictions by densely connected fusion layer.
In (15), the results of deep and fuzzy logic depiction part are represented by o d and o f are combined to the weight w d and w f , correspondingly. Next, the combined data are deeply converted via positioning multiple all connecting layers succeeding the fusion layer. The outcome integrates fuzzy degree and neural depictions completely when there is no more fuzzy degree. Though alternate feature extraction techniques for deep fusion are available, the fuzzy learning because of succeeding three reasons. Initially, fuzzy learning gives a comfortable manner for decreasing the uncertainty among the input data. This significant uncertainty decrease search is a crucial characteristic of fuzzy method and hard to replace through additional learning methods. Next, fuzzy learning certainly results in soft logistic value (i.e., fuzzy depictions) in the interval of 0; 1 ð Þ. Furthermore, all dimensions in the neural outcome are to the extent of 0; 1 ð Þ (Eq. (14)). The amount of fusion and neural results is to same extent, resulting in that these 2 results are simply combined in the fusion part. Thirdly, fuzzy learning part permits task-driven variable learning process. Now, the exhausted hand-crafted variable tuning step undergoes replacement via smart data driven learning by backpropagation (BP).
Part IV-Task-driven part: The last part is the classifier layer which allocates the combined depiction to its respective classification. The soft-max function is employed to categorize the data point into distinct classes. Let f i ; y i ð Þ represent ith input and respective labels and p H f i ð Þ represents feed-forward conversion of FDNN from the input layer to the previous task-driven layer (i.e., red layer). Subsequently, the succeeding soft-max function utilizing output layer by cth entry is estimated as: where w c and b c denote regression coefficient and bias of cth classes andŷ i ¼ŷ i1 ; . . .;ŷ ik ½ , represent predictive label of NN by k class. Later, the mean square loss of FDNN is determined on m trained instances,

Experimental validation
For validating the efficiency of the AIBDA-ICH model, a series of experiments were performed using test ICH dataset and is simulated using Python 3.6.5 tool with additional packages. Figure 4  Simultaneously, on the applied fold-9, the AIBDA-ICH method has reached a sens. of 94.88%, spec. of 98.45%, precision of 96.55%, and acc. of 96.56%. At last, on the applied fold-10, the AIBDA-ICH methodology has   Figure 6 illustrates the ROC curve analysis of the AIBDA-ICH model under 10 folds. In order to comparative analysis, a group of methods employed for comparison are GC-SDL [29], U-Net [30], watershed algorithm with ANN (WA-ANN) [31], ResNexT [32], window estimator module to a deep convolutional neural network (WEM-DCNN) [33], CNN and SVM techniques. Figure 7 examines the study of ICH diagnoses results of AIBDA-ICH method using present techniques with respect to specificity and sensitivity. The figure exhibited the WA-ANN has emerged as minimum performance with the least sensitivity of 60.18% and specificity of 70.13%. Simultaneously, the U-Net technique has illustrated minimum optimum efficiency with WA-ANN technique over sensitivity of 63.10% and specificity of 88.60%. In addition, the SVM method has attained to modest sensitivity of 76.38% and specificity of 79.41%. Likewise, the WEM-DCNN method has shown optimum outcomes over sensitivity of 83.33% and specificity of 97.48%. Furthermore, the CNN method has attempted to display certain moderate outcomes over the sensitivity of 87.06% and specificity of 88.08%. Moreover, the ResNexT method has shown around higher outcomes on sensitivity of 88.75% and specificity of 90.42%. Similarly, the GC-SDL approach has demonstrated the related to prior technique greater outcome over sensitivity of 94.01%and specificity of 97.78%. Lastly, the presented AIBDA-ICH technique has attained maximum sensitivity of 94.96% and specificity of 98.59%.
The figure portrays the investigation of ICH diagnoses results of AIBDA-ICH method with present approaches with respect to precision and accuracy. The figure illustrated the WA-ANN has shown up the minimal performer over least precision and accuracy of 70.08% and 69.78%. Simultaneously, the SVM technique has established slight optimal performance with WA-ANN technique over precision and accuracy of 77.53% and 77.32%. Similarly, the CNN method has achieved a moderate precision and accuracy of 87.98% and 87.56%. Forward by, the U-Net approach has showcased optimum outcomes with precision and accuracy of 88.19% and 87%. Furthermore, the WEM-DCNN method has tried to display certain moderate outcomes with precision and accuracy of 89.9% and 88.35%. Moreover, the ResNexT method has represented around optimum outcomes with precision and accuracy of 95.2% and 89.320%. Similarly, the GC-SDL method has illustrated the compared to previous method maximum outcomes with precision and accuracy of 95.79% and 95.73%. At last, the projected AIBDA-ICH technique has attained the maximum precision and accuracy of 94.96% and 98.59%.

Conclusion
This paper has presented an effective AIBDA-ICH-based e-diagnosis model using CT images. The presented model utilizes IoMT devices for data acquisition process. The IoMT devices transfer the data to the cloud server for Fig. 7 Comparative analysis of AIBDA-ICH model with existing methods remote diagnosis. In addition, to handle big data, Hadoop ecosystem is utilized. Followed by, graph cut-based segmentation model is implemented for identifying the affected brain regions in the CT scan. Next to that, CapsNet model is employed to produce a proper set of feature vectors from the segmented image. Finally, FDNN model is applied to assign proper class labels to the input test images. Therefore, the application of AI and DL models aid to successfully diagnose the ICH of remote patients.
The experimental values pointed out the improved e-diagnostic performance of the AIBDA-ICH model over the other compared methods with the precision and accuracy of 94.96% and 98.59%. As a part of future scope, the performance of the AIBDA-ICH model can be improvised by the use of hyperparameter tuning strategies. Besides, new ICH diagnostic model with DL-based segmentation technique can be employed to improve the overall performance of the proposed model.