Sage, A. Badura, P introduced a model that is based on double branch convolutional neural network Resnet-50 architecture to identify the various types of intracranial haemorrhage. Support vector machine and RF were the two classifiers tested for classification (random forest). The model, which had an accuracy of 93.3% for intraparenchymal haemorrhage and 96.7% for intraventricular haemorrhage, was best suited to random forest.[1]
Anupama C.S.S., Sivaram M., Lydia, proposed the GC-SDL model named as (Grab Cut based synergic deep learning model). First the Gabor filter is used to remove the noise and the grab cut segmentation process is used to locate the unhealthy position in the images. Then feature extraction is performed using Synergic DEEP LEARNING model, and the SoftMax layer is used as a classifier. The GC-SDL model achieved 97,78% specificity, 95.79% precision, and 95.73% accuracy.[2]
Brain CT images' identification Awwal Dawud, Muhammad & Yurtkan Kamil & Oztoprak Huseyin uses the CNN to approach a deep learning technique. SVM classifier is used with a modified unique version of Alex Net.[3]
For the purpose of precise prediction of cerebral haemorrhage, B. Nageswara Rao presented an automatic transfer deep learning system combining ResNet-50 and dense layer. The proposed method categorizes each individual CT images as either hemorrhagic or normal based on the input data given. This deep transfer learning approach achieved an accuracy of 99.6%, 99.7% specificity, and 99,4% sensitivity comparing with Resnet-50.[4]
Venugopal D, introduced deep learning to develop a multi-modal data fusion-based feature extraction technique (FFEDL). initially gaussian filtering is used to the images, then segmentation is done using density-based fuzzy c means and feature extraction is performed using Reset-152 architecture.[5]
Sundar Santhoshkumar & Varadarajan Vijayakumar & Gavaskar S. & Amalraj J. & Sumathi A., proposed a model for cutting-edge image processing, using the DN-ELM system named as (Dense Net with extreme learning machine). This is a combination model of Tsallis entropy and Grasshopper Optimisation Algorithm (TE-GOA) and then dense Net is used to extract the features of segmented images. The photos are finally classified using ELM. Maximum accuracy provided by this model was 96.34%.[6]
Ye H., Gao F., Yin Y. suggested a combination CNN-RNN model for the purpose of identifying cerebral bleeding and its five subtypes. By analysing the 3D CT scans, the system determined whether bleeding would occur within 30 seconds.[7]
A system built on the architecture of deep neural networks was proposed by Burduja Mikhail & Ionescu Radu Tudor & Verga Nicolae. It is made up of CNN, which will use LSTM (long short term memory) for effective processing after receiving the CT images as input. Grad-CAM will be incorporated into this system's future visualisation in order to provide a full explanation for the prediction and to make it ready for situations where quick diagnosis is required.[8]
Duc Tong Phong & Hieu Duong & Nguyen Hien & Nguyen Trong & Nguyen Hong Vu & Tran Hoa et.all used a deep learning to create a system for identifying brain haemorrhages. They experimented with LeNet, GoogLeNet, and inception-ResNet, three different forms of convolutional neural networks. With Google LeNet and Inception ResNet, they merely trained the final fully linked layer, whereas LeNet required training for all layers. They believed that GoogLeNet and inception-ResNet may be utilised for medical picture identification, particularly for brain haemorrhage, and they validated that LeNet is the model that takes the longest to train.[9]
Six machine learning models, that includes decision trees, random forests, gradient boosted trees, K-Nearest Neighbor, support vector machines, and logistic regression, were proposed by Goyal Rushank. 160 CT images with and without utility matrices are trained here, which is a form of cost-sensitive learning. Therefore. SVM outperformed all the others with accuracy rates of 97.5% without penalty and 92.5% with penalty.[10]
Without CNN, the total number of CT images taken for each patient, or the zone below ROC curve, Lee Ji & Kim Jong & Kim Tae & Kim Young Soo. proposed a model that can recognize ICH. (AUC). Based on intracranial height, CT images were segmented in to 10 units for the purpose of ICH localization. The subtype of subarachnoid haemorrhage is where this test performs best. (SAH).[11]
The proposed technique by Shanmuganathan Sankar & Nivedhitha & Ranganayakulu Dhanalakshmi., uses CT images to locate the haemorrhage. To prepare them for further processing, the photos go through pre-processing. Initially the images are pre-processed with grey scale conversion, scaling the images, detect the edges and sharpening. After pre-processing, the images go through morphological processes that help determine the traits that correspond to the haemorrhage’s shape. Closure reconstruction, Sobel, and markers are utilised to emphasise the relevant area in the processed CT image. The image is subsequently subjected to the watershed segmentation method. The suggested system divides haemorrhages into three categories. The system's classification accuracy for the three forms of haemorrhage was found to be 98% on average.[15]
Ammar Mohammed & Lamri Mohamed & Saïd Mahmoudi & Laidi Amel., designed a tool to help radiologists to identify Intracranial Haemorrhage and each of their subtypes by diagnosing the CT scan images. Five kind of deep learning models: ResNet 50, VGG16, Xception, InceptionV3, InceptionResnetV2 had been used here. The result showed that VGG16 architecture had an accuracy of 96%.[14]
A deep learning model was used by Mushtaq Muhammad & Shahroz Mobeen & Aseere Ali & Shah Habib & Majeed Rizwan & Shehzad Danish & Samad Ali. to categorize the brain haemorrhages. Here, hybrid CNN models of CNN + LSTM and CNN + GRU were applied. The purpose of this study is to develop a unique CNN architecture called Brain Haemorrhage Classification using neural networks. (BHCNet).[13]
Chen Hang & Khan Sulaiman & Kou Bo & Nazir Shah & Liu Wei & Hussain Anwar. identified brain haemorrhages using an intelligent IoT-based solution with machine learning techniques. A combination model of a support vector machine and a feedforward neural network was applied to the CT scan images of the intracranial haemorrhage. The classification rates for SVM and FNN are 80.67% and 86.7%, respectively. It is found that the FNN outperforms the backward neural network for the identification of intracranial images.[12]
An unsupervised PCA-Net was proposed by Ganeshkumar M. & Vishvanathan Sowmya & Gopalakrishnan E. A & Soman K. to excerpt features from CT images. (Principal Component Analysis). Additionally, the K-means classifier is trained to recognize ICH in an entirely unsupervised manner, without the need for any class labels, using the features gathered from PCA-Net. A supervised linear SVM classifier was further trained using the extracted PCA-Net features. All three models - K-means, PCA-Net, and linear SVM were trained entirely unsupervised. The weighted average accuracy, precision, recall, and weighted average Fl-score of the proposed method were all 67%.[16]
Jangaraj Avanija & Sunitha Gurram & Madhavi Reddy & Vittal R., proposed an early cerebral haemorrhage detection by utilizing an automated model and DenseNet. DenseNet is used to analyze MRI images to find cerebral haemorrhage and its various subtypes. Deep linked convolution networks, or DenseNet, are another name for them. The cerebral haemorrhage variation is predicted based on the image segments. The proposed approach for detecting cerebral haemorrhage using convolution neural network is 91% achieved using the gradient from loss function, however DenseNet layers are much narrower and add a smaller number of feature maps, performing better in comparison.[17]
Alfaer Nada & Aljohani Hassan & Abdel-Khalek Sayed & Alghamdi Abdulaziz & Mansour Romany proposed an automatic intracerebral haemorrhage detection which is Fusion Based Deep Learning With Swarm Intelligence named as AICH-FDLSI. Pre-processing, picture segmentation, feature extraction, and classification are its four processes. The noise from the images is first removed using median filtering, after which the images are segmented, followed by the features extraction and hyper-parameter optimization processes, and finally by FSVM classification.[18]
Asqiriba Hassan & Sultan Gazi, this paper includes about some optimal image segmentation techniques. Aim of this report was to explore and present an optimal and efficient image segmentation method.[19]