Segmentation and Approximation of Blood Volume in Intracranial Hemorrhage Patients Based on Computed Tomography Scan Images Using Deep
Learning Method
1Kezia Irene, 1M. Anwar Ma’sum, 2Reyhan Eddy Yunus, and 1Wisnu Jatmiko
1Faculty of Computer Science Universitas Indonesia Kampus Baru UI Depok – 16424 Jawa Barat, Indonesia
2Department of Radiology, Faculty of Medicine, Universitas Indonesia, Jakarta 10430, Indonesia Email: [email protected]
Abstract—Traumatic brain injury is a common injury that can range from mild concussions to severe permanent brain damage.
One of the severe damages caused by traumatic brain injury is intracranial hemorrhage, which is typically diagnosed by clinicians using head computed tomography (CT) scans. However, in some hospitals in Indonesia, sometimes there is a lack of clinicians who are able to interpret the CT scan results, leading to morbidity and mortality. Deep learning algorithms, especially convolutional neural networks (CNN) can be utilized to help clinicians in diagnosing patients with intracranial hemorrhage. In this study, we propose an automated segmentation and blood volume approximation of intracranial hemorrhage patients from CT scan images using deep learning and regression methods. For the blood segmentation, we utilized Dynamic Graph Convolutional Neural Network (DGCNN) architecture and for the blood volume approximation, we utilized regression methods. The dataset for this work consists of 27 head CT scans obtained from the Cipto Mangunkusumo National General Hospital 2019 traumatic brain injury data segmented manually by a radiologist.
For blood segmentation, we proposed several scenarios by upsampling or downsampling the data. The best results obtained in the scenario without doing upsampling resulted in a sensitivity of 97.8% and a specificity of 95.6%. For blood volume approximation, the best results are obtained using the support vector machine (SVM) method with a radial basis function (RBF) kernel, with a mean squared error of 3.67x10^4.
Keywords—Intracranial hemorrhage, CT scan, three- dimensional CNN, DGCNN, blood volume approximation
I. INTRODUCTION
Traumatic brain injury is a physical injury to the brain tissue that temporarily or permanently damages the brain function [1].
The number of people suffering from this injury in developing countries increases significantly every year [2]. Traumatic brain injury in developing countries is generally caused by traffic accidents and generally occurs in young patients [3]. One of the diseases that can be caused by head trauma is intracranial bleeding. Intracranial hemorrhage refers to all bleeding within the intracranial space, including the brain parenchyma and the surrounding meningeal space [4]. Measurement of blood volume in patients with intracranial hemorrhage is important because it will determine further treatment [7].
Attempt to treat intracranial hemorrhage is done by first examining the patient's brain by taking computed tomography (CT) images [1]. CT is a method of taking pictures of internal organs that is widely used to detect intracranial hemorrhage [5].
CT works by utilizing different absorbencies in body tissue when placed between the transmitter and the x-ray detector. On CT images, the absorption of body tissues such as the brain, blood, muscle, and bone tissue are characterized by differences in Hounsfield units (HU). Intracranial hemorrhage is characterized by the presence of blood, which has higher HU than brain tissue but is lower than bone [6].
Currently, in Indonesia, CT images of patients will be examined by a radiologist, which may lead to human error or late diagnosis which results in a high mortality rate for intracranial hemorrhage patients. Making automated segmentation and determining the blood volume from CT scan images of the intracranial hemorrhage patients can help clinicians to treat patients more quickly, reducing the mortality rate.
In detecting and classifying CT scan data, various machine learning methods have been used. One of the most used methods is the two-dimension convolutional neural network.
This method is done by changing the classification of three- dimensional objects into two dimensions. Some studies have successfully produced a sensitivity of more than 90% by using this method [8]. However, there are drawbacks to the two- dimensional convolutional neural network method; the potential loss of spatial contextual information when volumes are analyzed using two-dimensional slices [9].
A three-dimensional convolutional neural network (3D CNN) is an emerging architecture that is used widely to perform video analysis or volumetric medical images [10]. In the analysis of medical images, 3D CNN has been used in detecting abnormalities in CT images and magnetic resonance imaging (MRI) [18]. Dou et al. [11] used the 3D CNN method to detect cerebral microbleeds on MRI brain images and produce 93%
sensitivity, although it has a precision of only 44%.
One of the 3D CNN methods is to make the segmentation on point cloud data. In 2017, Qi et al. [15] developed PointNet that made it possible to segment and classify point cloud data directly. Wang et al. [14] developed the Dynamic Graph CNN (DGCNN) whose architecture is the development of PointNet [15]. The DGCNN model increases the accuracy of semantic segmentation to 84.1% compared to the 78.5% accuracy of the PointNet model. The prediction results from the DGCNN model are in the form of objects consisting of x, y, z data, color, and also predictions that facilitate the reconstruction and approximation of the predicted object volume.
Based on these problems, this research is conducted to obtain a prediction of intracranial hemorrhage based on CT scan images using deep learning methods with DGCNN architecture.
This study also aims to estimate the volume of bleeding based on prediction results at the deep learning stage. After that, the results obtained from deep learning segmentation and volume approximation will be compared with the segmentation and volume calculation that has been made by the radiologist.
II. RELATEDWORK
Research related to intracranial hemorrhage segmentation has been carried out with a variety of methods, ranging from only using image processing to using deep learning methods.
Research related to intracranial hemorrhage segmentation was first carried out with the supervised learning approach by Chan, T [16] who proposed the method of segmentation of intracranial hemorrhage by combining and making parallel all slices of CT head images. After that, the candidate areas for intracranial hemorrhage are selected using top-hat transformation and extraction of areas with high-intensity levels. Finally, the candidate area is included in the knowledge- based classifier for intracranial hemorrhage segmentation. This method produces 100% sensitivity in the slice level, 84.1%
specificity in the slice level, and 82.6% sensitivity in the lesion- level. This approach was also carried out by Muschelli et al.
[17] by extracting the brain using a CT brain-extracted template, after which feature extraction was performed from each CT head slice. These features consist of image intensity information, average, and standard deviations, images of contralateral differences from the template, distance to the center of the brain, and differences in intensity with the normal brain. After that, the classification model is created using logistic regression methods, generalized additive models, and random forests. These models were trained on 10 CT scans and tested on 102 CT scans, resulting in the highest Dice coefficient of 0.899.
Several other studies have conducted an unsupervised learning approach and pre-processing CT head images to eliminate skull and noise on CT images [18-20]. These studies are based on unsupervised clustering to make segmentation in the bleeding section. Prakash et al.,[18] and Shahangian et al., [20] used the Distance Set Evolution (DRLSE) method to adjust active contours in intracranial hemorrhage areas. The first step in their method is segmentation of the brain by eliminating the skull. Next, an intracranial hemorrhage segmentation was performed based on the DRLSE method. Then, extraction of the shape and texture of the intracranial hemorrhage feature was performed. And finally, detection of intracranial bleeding. This method produces a Dice coefficient of 58.5, a sensitivity of 82.5%, and a specificity of 90.5% on 627 CT slices.
Research that approaches using deep learning for segmentation of intracranial hemorrhage, some using the CNN method [21-22] and others [23-24] use the fully convolutional networks (FCN). Chang, P., et al. [22] developed a deep learning algorithm to detect, segment intracranial hemorrhage, and calculate the volume of bleeding. The deep-learning model used is a development of the CNN region-of-interest that estimates areas that are blood for each slice of the CT image, then produces segments based on these regions. The results obtained were a sensitivity of 95%, a specificity of 97%, 0.97 area under the curve (AUC), and an average dice score of 0.85 for bleeding segmentation. While Hssayeni, Murtadha used the FCN approach on the U-Net architecture which produced a Dice coefficient of 0.31, 97.28% sensitivity, and 50.4%
specificity.
In recent years, segmentation using point clouds has been applied for medical image analysis with promising results. F.
Zanjani et., al. [28] applied image segmentation on intra-oral scans (IOS) achieving 0.94 IOU score. A. Sekuboyina, et. al, [29] used point clouds for vertebra shape analysis, resulting in area-under-ROC curve of >95%. F. Drokin, I., & Ericheva, E, [30] applied the DGCNN method on chest CT scan images to make the segmentation of lesions, resulting in 85.98 free- response roc curve (FROC).
From these studies, we decided to utilize point clouds with DGCNN method to make the segmentation for head CT scans in patients with intracranial hemorrhage. This method is chosen because of the promising results of previous studies using point clouds in medical images. Using point clouds data is also suitable for medical images because point clouds is able to store multiple data in each point so that there will be no missing information from the CT scan images.
III. RESEARCHMETHODOLOGY A. Data Gathering
We obtained data through third parties in the form of CT image data of head trauma patients in 2019 from Cipto Mangunkusumo Hospital. The data obtained in the form of DICOM data that has been annotated by a radiologist. There were 27 CT images of the head of an intracranial hemorrhage patient who had been annotated by a radiologist. CT slices of the head can be seen in Fig. 1.
Fig. 1. CT Head Slices
CT images consist of 30 to 50 cross sections that are at different levels. The head CT data obtained has a thickness of 5 millimeters with each image consisting of 512 x 512 pixels.
Each CT data also has the name, identification number, gender, and date of the CT image taken of the patient. Summary of data on intracranial hemorrhage patients can be seen in Table I.
TABLE III. INTRACRANIAL HEMORRHAGE PATIENTS’DATA
Number of Patient 27 Slices per patient 30 to 50
Slice thickness 5 mm
Pixel number per slice 512 x 512
B. Implementation
Implementation is done to identify bleeding and estimate blood volume from CT images of the head. Implementation begins with pre-processing of the head CT image, followed by segmenting the bleeding with DGCNN, then calculating the bleeding volume from the results of the bleeding segmentation using 3D volume rendering. Implementation is carried out in stages according to Fig. 2.
C. Data Preprocessing
The first step taken in the dicom dataset is data augmentation. This was done based on the recommendation of head data processing by Muschelli, J. [12] aimed at multiplying training objects and patching model generalizations. For each head piece, a rotation of 180 degrees, flip, and flip and rotation are performed so that one image is segmented into four images.
Fig. 3 illustrates an example of a head CT image section that is segmented into four CT head images. After the data augmentation process, the total CT images of the patient's head became 108 images
Fig. 3. CT Head Augmentation
After augmentation of the dataset, each image is processed to remove empty air and skull bones by windowing the Hounsfield Unit (HU) from the CT head image. HU air is -1000 and based on Birur, et. al. [27], the bone HU was more than 300.
Based on these references, on the CT image dataset, HU windowing was performed between -999 to 299. Figure 4 shows the results of windowing on the CT head image.
Fig. 4. HU Windowing
After windowing the head CT image, it can be seen in Figure 5, that the non-blood portion of the brain image has a much wider area than the bleeding area. This causes the number of class 1 points (blood) is much less than the class 0 point (non- blood) which will complicate the deep learning process. Based on Muschelli, J. [12], down sampling needs to be done in class 0 so that the number of class 0 and class 1 approaches each other as in Fig. 5.
Fig. 2. Implementation Diagram
Fig. 5. Downsampling class nonblood
The down sampling process is done by taking all the class 1 points and then counting the number of class 1 points. After that, the points of class 0 are taken from a number of class 1 ceilings by dividing the average points taken at each level of CT images. In each level, a random number generator is used to get any points that need to be taken.
After the down sampling process, 4096 points were sampled in each block for data training purposes [15]. Sampling data in the form of .h5 files that can be read directly when doing training on the model.
D. Dynamic Graph Convolutional Neural Network (DGCNN) Dynamic Graph Convolutional Neural Network (DGCNN) is a deep learning architecture that utilizes the concept of convolutional neural networks and point clouds. DGCNN is a development of PointNet, with renewal in the form of Edge Convolution (EdgeConv). EdgeConv is a convolution layer that is useful for taking local geometric structures while maintaining permutation invariance [14].
Fig. 6. EdgeConv in DGCNN
In Fig. 6, EdgeConv takes local geometric structures by constructing graphs at adjacent points and applying convolution operations on each connected edge [14].
In the DGCNN architecture, there are four EdgeConv layers to extract geometric features from each point and there is also a pooling layer like PointNet to get global features. The DGCNN classification produces an output of N points along with a classification value for C class.
Fig. 7. DGCNN Segmentation Architecture
Fig. 7 shows DGCNN segmentation architecture. Segmentation architecture has a different number of EdgeConv layers in a ratio of 2: 3. In the segmentation architecture, there is a shortcut connection that is used to enter all the output from the EdgeConv layer to describe local features. The output of DGCNN segmentation is in the form of N points along with P classification probability for each segment for each point.
Training on DGCNN is carried out using a 4-fold cross validation strategy with 81 images used as training data and 27 images used as testing data. The input data used when conducting training is in the form of point cloud data in HDF5 format consisting of the pixel value for each slice (x and y), the level of the slice (z), the HU value, and blood or non-blood labels. Testing data is randomly divided so that it is done once testing for all images. Data not taken as testing data is used as training data. The model configuration in the training is explained in Table II as follows:
TABLE IV. DGCNNMODEL CONFIGURATION
Configuration Values
Epoch 50
Batch Size 12
Learning Rate 0,001
Optimizer Adam/Momentum
Momentum 0,9
The configuration used is the default configuration on the DGCNN model [14]. The author made no changes to the DGCNN architecture used by Wang et al. [14].
E. Semantic Segmentation
In blood detection, the separation between blood points and non-blood points is done by semantic segmentation of the data.
Semantic segmentation is done using a model of the results of training on the data of each image. Semantic segmentation results in the form of class predictions at a point, so that predictions on an image are a combination of prediction results at each point.
The output of the semantic segmentation is in the form of a text file with an explanation of the x, y, z values, and also the bleeding prediction for each point on the patient's head CT image. In evaluating the output results, each point of the prediction result will be compared with the ground truth points which are the result of manual segmentation by the radiologist.
F. Segmentation Visualization
Three-dimensional visualization of semantic segmentation is done using Meshlab software. After that, a mapping of semantic segmentation results into a dicom file to facilitate visualization for radiologists. The mapping process is done by changing the metadata from the dicom file by using the points resulting from semantic segmentation.
G. Linear Regression
Linear regression is a linear approach to modeling the relationship between a scalar response (or dependent variable) and one or more explanatory variables (or independent variables). Linear regression has a purpose to evaluate the relative impact of the predictor variables on a particular outcome. The simple linear regression model contains only one independent variable (explanatory), Xi for i=1,...,n subjects, and is linear with respect to the regression parameters and interdependent variables. The appropriate dependent variable (result) is labeled. The linear regression model is expressed as
𝑌𝑖 = 𝑎 + 𝑏𝑋𝑖+ 𝑒𝑖 (1)
where the regression parameter a is the addition (on the y axis), and the regression parameter b is the gradient on the regression line. Random error 𝑒𝑖 are assumed to be uncorrelated, with an average of 0 and a constant variance. For convenience in inference and increased efficiency in estimation [26], analysis often raises additional assumptions that errors are normally distributed.
The steps in analyzing the regression model are: determine whether the assumptions underlying the relationships between variables are fulfilled in the data, find the equation that best matches the data, evaluate the equation to determine the strength of the relationship for predictions and estimates, and assess whether the data are in accordance with the criteria - previous criteria, before the equation is applied to predictions and estimates.
H. Support Vector Machine
Support Vector Machine (SVM) was developed by Vapnik and Lerner [25] as a machine learning model that classifies and regresses by creating hyperplane with high or unlimited dimensions. Unlike linear regression, SVM maps initial parameter vectors into feature space with higher dimensions through nonlinear kernel functions. Without the need to explicitly calculate nonlinear mapping, point products can be calculated efficiently in higher dimensional spaces. The dominant feature that makes SVM very interesting is that classes that are separated nonlinearly in the original space can be linearly separated in higher dimensional feature spaces. Thus SVM is able to solve complex nonlinear classification problems.
I. Blood Volume Measurement
Calculation of blood volume is estimated using linear regression, SVM with linear kernel, and SVM with RBF kernel.
The features used are the predictive results points of the bleeding segmentation and basic truth, i.e. the results of the segmentation of the radiologist. The linear methods were chosen because they are suitable for data calculations and small variables. After going through a regression, the calculation is done by calculating the mean squared error, the coefficient of determination, and the mean absolute error.
IV. RESULTSANDANALYSIS A. Evaluation Method
An evaluation is carried out to measure the results of implementation in identifying bleeding. An evaluation is also conducted to determine the performance of the model at the semantic segmentation stage and the bleeding labeling stage.
In measuring the statistical accuracy of medical data, there are several measurement methods that are often used, namely sensitivity, specificity, dice coefficient, and accuracy.
However, in this study, several methods of measuring the results of other experiments will also be used, namely precision and intersection over union (IoU).
Evaluations are carried out to measure the results of implementation in approximating bleeding volume using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (RMSE).
B. Training Scenarios
Training on DGCNN will be carried out using a 4-fold cross validation strategy with 81 images used as training data and 27 images used as testing data. The division of data in the testing data is randomly divided so that all images must have been tested once. Data not taken as testing data is used as training data.
For bleeding segmentation testing, there are 3 scenarios created.
These scenarios are created by doing up sampling on class 1 (blood) data. Following are the test scenarios in the research implementation:
1. Scenario A, semantic segmentation is done without up sampling on class 1 data
2. Scenario B, semantic segmentation is carried out by sampling up to class 1 data a number of times the initial data
3. Scenario C, semantic segmentation is carried out by sampling up to class 1 data 1.5 times the initial data The volume calculation will be done using several regression methods and the results will be compared. The methods used are linear regression methods, support vector machines - radial basis functions, tree regression, and random forest regression.
C. Semantic Segmentation of CT Head Dataset
The DGCNN architecture utilized to make blood segmentation is the model generated at epoch 50. Table III explains the evaluation results on the whole picture using the accuracy, recall, precision, f-Score and IoU evaluation metrics.
The results of blood evaluation are indicated by the variable B (blood) and NB (not blood) for non-blood objects.
Based on the evaluation results in Table III, it can be seen that the Momentum optimizer has a better level of accuracy than Adam's optimizer. It can also be seen that data that is not up- sampled has better accuracy, sensitivity, and specificity than the data done by up-sample.
This is because the training model using up sampling has class 1 (blood) points that are denser than without using up sampling. This causes the deep learning model to learn if a point is more closely located, it will be classified into blood classes.
This conclusion is also supported by the data in Table III, which can be seen that the precision, recall, F-score, and IoU for the blood class are lower than the nonblood class. Therefore, the segmentation results produced by the Momentum optimizer without an up sample are used to calculate the bleeding volume and are processed into visualizations on CT images.
D. Visualization of Semantic Segmentation Results of CT Head Dataset
After segmenting the head CT image, an .obj extension file is generated that can be seen visually and three dimensions through the Meshlab software. The results of the visualization produced can be seen in Fig. 8. Green dots denote not blood while dark blue dots denote blood.
Fig. 8. Segmentation Visualization Result
In order to facilitate the radiology doctor in seeing the visualization of bleeding, a mapping of segmentation results is made into the dicom image. Fig. 9 explains the results from mapping the results of segmentation compared to ground truth which is the result of the segmentation of the radiologist.
(a) (b)
Fig. 9. Left: Ground truth brain CT scan, Right: Semantic Segmentation Result
E. Blood Volume Approximation Testing
The results of blood volume calculations are done using linear regression, linear SVM, and SVM with RBF kernel can be seen in Fig. 10. These methods are used to approximate the blood volume from the semantic segmentation result because there is no definite way to convert from point cloud to blood volume in cm3 unit. When we tried to use a constant to multiply the point cloud into the cm3 unit, we discovered that there are differences in density within each patients’ data.
(a)
(b) TABLE III. SEMANTIC SEGMENTATION EVALUATION
(c)
Fig. 10. a: Linear Regression Result, b.: SVM Linear Result, c.: SVM – RBF Result
Evaluation of blood volume approximation can be seen in Table IV. Based on the evaluation results, it can be seen that SVM with RBF kernel has the lowest mean squared error and the lowest root mean squared error among the three methods used. However, linear SVM has the lowest mean absolute error and the lowest mean absolute percentage error among the three methods used.
TABLE IV. BLOOD VOLUME APPROXIMATION EVALUATION
Metode Mean
squared error
Root Mean squared
error
Mean Absolute
Error
Mean Absolute Percenta ge Error Linear
Regression
7.33 x 1010 2.71 x 105 1.88 x 104 187.66
SVM - Linear 4.99 x 1011
7.07 x 105 1.08 x 104 107.78 SVM - RBF 1.34 x 𝟏𝟎𝟗 3.67 x 𝟏𝟎𝟒 9.99 x 𝟏𝟎𝟑 99.95
Overall, the three methods used produce mean squared error, root mean squared error, and mean absolute error more than 103. A fairly high root mean squared error states that there are some data that are outliers. While the mean absolute error is high enough to indicate data that has high variation so that the function taken cannot make generalizations from existing features. The cause of the high overall evaluation metric results is due to the lack of data in the experiment and the data entered in the results of semantic segmentation that are not yet truly accurate.
V. CONCLUSION
In this study, the process of prediction and segmentation of bleeding is done using dataset of CT heads of patients with intracranial hemorrhage and also the estimated volume of bleed.
The prediction process starts by preprocessing the CT head image, then proceeded by semantic segmentation and visualizing the results of segmentation. After that, the volume calculation is done by doing regression between the results of segmentation and the ground truth.
Blood segmentation on CT images of intracranial patient's head can be done with the deep learning method using DGCNN architecture. The steps taken in predicting intracranial
hemorrhage begins with the pre-processing of the image, then semantic segmentation using the deep learning method, and continues with the visualization process on the CT head image.
The best bleeding segmentation is done using a momentum optimizer without up sampling reaching 98% sensitivity.
Approximate bleeding volume from CT head images with the best accuracy can be obtained by regression using support vector machine (SVM) with RBF kernel with the points of prediction results of semantic segmentation and ground truth reaching the mean absolute percentage error of 99.95. The use of DGCNN in predicting intracranial hemorrhage with small datasets has a promising result but it would be better to have more datasets.
VI. FUTURE WORK
We realize that the final results of segmentation and approximation of intracranial hemorrhage volume are far from perfect and still have suboptimal performance, therefore the authors have some suggestions regarding this study. We recommend using a larger number of datasets to add validation to the segmentation of intracranial hemorrhagic head CT images. We also recommend experimenting with machines with large GPUs and memory in order to work on larger datasets more quickly.
We emphasize that the results of the experiments conducted by the author cannot be generalized to segment and approximate the volume of intracranial bleeding because the dataset used is still limited and validation data that is considered ground truth can vary between radiologists.
Therefore, the authors suggest increasing data validation by cross-checking with different radiologists.
Next, we suggest evaluating the detection of intracranial bleeding using other metrics such as Intersection over Union and Mean Average Precision in order to obtain a more in- depth analysis of the research results. We have problems using these metrics due to the limited number of datasets and the implementation of metric calculations on object detection data in the form of point clouds.
ACKNOWLEDGMENT
This work is supported by the grant from Universitas Indonesia entitled "Hibah Publikasi Terindeks Internasional (PUTI)" with No: NKB-2145/UN2.RST/HKP.05.00/2020 and Faculty of Medicine, Universitas Indonesia.
REFERENCES
[1] Parikh, S., Koch, M., & Narayan, R. K. (2007). Traumatic Brain Injury.
[2] Maas, A. I., Stocchetti, N., & Bullock, R. (2008). Moderate and severe traumatic brain injury in adults.
[3] Finfer SR, Cohen J. Severe traumatic brain injury.
Resuscitation 2001; 48: 77–90
[4] Caceres, J. A., & Goldstein, J. N. (2012). Intracranial hemorrhage. Emergency medicine clinics of North America, 30(3), 771.
[5] Heit JJ, Iv M, Wintermark M (2017) Imaging of intracranial hemorrhage. J Stroke 19:11–27.
[6] Carney N, Totten AM, O’Reilly C et al (2017) Guidelines for the management of severe traumatic brain injury, Fourth Edition. Neurosurgery 80(1):6–15.
[7] Christoforidis, G. A., Slivka, A., Mohammad, Y., Karakasis, C., Avutu, B., & Yang, M. (2007). Size matters:
hemorrhage volume as an objective measure to define significant intracranial hemorrhage associated with thrombolysis. Stroke, 38(6), 1799-1804.
[8] Roth, H.R.; Lu, L.; Seff, A.; Cherry, K.M.; Hoffman, J.;
Wang, S.; Liu, J.; Turkbey, E.; Summers, R.M. A New 2.5 D Representation for Lymph Node Detection Using Random Sets of Deep Convolutional Neural Network Observations. In Proceedings of the International Conference on Medical Image Computing and Computer- Assisted Intervention, Boston, MA, USA, 14–18 September 2014; pp. 520–527
[9] Ker, J., Singh, S. P., Bai, Y., Rao, J., Lim, T., & Wang, L.
(2019). Image thresholding improves 3-dimensional convolutional neural network diagnosis of different acute brain hemorrhages on computed tomography scans.
Sensors, 19(9), 2167.
[10] Moeskops, P.; Viergever, M.A.; Mendrik, A.M.; de Vries, L.S.; Benders, M.J.; Išgum, I. Automatic segmentation of MR brain images with a convolutional neural network.
IEEE Trans. Med. Imaging 2016, 35, 1252–1261.
[11] Dou, Q.; Chen, H.; Yu, L.; Zhao, L.; Qin, J.; Wang, D.;
Mok, V.C.; Shi, L.; Heng, P.A. Automatic detection of cerebral microbleeds from MR images via 3D convolutional neural networks. IEEE Trans. Med. Imaging 2016, 35, 1182–1195
[12] Muschelli, J. (2019). Recommendations for Processing Head CT Data. Frontiers in neuroinformatics, 13, 61.
[13] Wang, Y., Sun, Y., Liu, Z., Sarma, S. E., Bronstein, M. M.,
& Solomon, J. M (2019). Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5), 1-12.
[14] Wang, Y., Sun, Y., Liu, Z., Sarma, S. E., Bronstein, M. M.,
& M.Solomon, J. (2018, January 24). Dynamic Graph CNN for Learning on Point Clouds. arxiv:
https://arxiv.org/pdf/1801.07829.pdf
[15] Qi, C. R., Su, H., Mo, K., & Guibas, L. J. (2017). PointNet:
Deep Learning on Point Sets for 3D Classification and Segmentation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (pp. 77-85). Honolulu.
[16] Chan, T. Computer aided detection of small acute intracranial hemorrhage on computer tomography of brain.
Computerized Medical Imaging and Graphics 2007, 31, 285–298.
[17] Muschelli, J.; Sweeney, E.M.; Ullman, N.L.; Vespa, P.;
Hanley, D.F.; Crainiceanu, C.M.:Primary intracranial hemorrhage probability estimation using random forests on CT. NeuroImage: Clinical 2017,14, 379–390. (Prakash, KB, et al., 2012, Bhadauria, H., et al., 2014, Shahangian, B. , et al., 2016, Gautam, A., et al., 2019)
[18] Prakash, K.B.; Zhou, S.; Morgan, T.C.; Hanley, D.F.;
Nowinski, W.L. Segmentation and quantification of intra- ventricular/cerebral hemorrhage in CT scans by modified distance regularized level set evolution technique. Int. J.
Comput. Assist. Radiol. Surg. 2012, 7, 785–798.
[19] Bhadauria, H.; Dewal, M. Intracranial hemorrhage detection using spatial fuzzy c-mean and region-based active contour on brain CT imaging. Signal Image Video Process. 2014, 8, 357–364.
[20] Shahangian, B., & Pourghassem, H. (2016). Automatic brain hemorrhage segmentation and classification algorithm based on weighted grayscale histogram feature in a hierarchical classification structure. Biocybernetics and Biomedical Engineering, 36(1), 217-232.
[21] Lee, H., Yune, S., Mansouri, M., Kim, M., Tajmir, S. H., Guerrier, C. E., ... & Gonzalez, R. G. (2019). An explainable deep-learning algorithm for the detection of acute intracranial haemorrhage from small datasets. Nature Biomedical Engineering, 3(3), 173
[22] Chang, P. D., Kuoy, E., Grinband, J., Weinberg, B. D., Thompson, M., Homo, R., ... & Filippi, C. G. (2018).
Hybrid 3D/2D convolutional neural network for hemorrhage evaluation on head CT. American Journal of Neuroradiology, 39(9), 1609-1616.
[23] Nag, M. K., Chatterjee, S., Sadhu, A. K., Chatterjee, J., &
Ghosh, N. (2019). Computer-assisted delineation of hematoma from CT volume using autoencoder and Chan Vese model. International journal of computer assisted radiology and surgery, 14(2), 259-269.FCN
[24] Kuo, W.; Häne, C.; Yuh, E.; Mukherjee, P.; Malik, J. Cost- Sensitive active learning for intracranial hemorrhage detection. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2018; Frangi, A.F., Schnabel, J.A., Davatzikos, C., Alberola-López, C., Fichtinger, G., Eds.; Springer International Publishing:
Cham, Switzerland, 2018; pp. 715–723.
[25] Smola, A. J., & Schölkopf, B. (2004). A tutorial on support vector regression. Statistics and computing, 14(3), 199- 222.
[26] Chen, C., Twycross, J., & Garibaldi, J. M. (2017). A new accuracy measure based on bounded relative error for time series forecasting. PloS one, 12(3).
[27] Birur, NPraveen; Patrick, Sanjana; Gurushanth, Keerthi;
Raghavan, AShubhasini; Gurudath, Shubha (2017).
"Comparison of gray values of cone-beam computed tomography with hounsfield units of multislice computed tomography: An in vitro study". Indian Journal of Dental Research.
[28] Zanjani, F. G., Moin, D. A., Verheij, B., Claessen, F., Cherici, T., & Tan, T. (2019, May). Deep learning approach to semantic segmentation in 3d point cloud intra- oral scans of teeth. In International Conference on Medical Imaging with Deep Learning (pp. 557-571).
[29] Sekuboyina, A., Rempfler, M., Valentinitsch, A., Loeffler, M., Kirschke, J. S., & Menze, B. H. (2019, October).
Probabilistic Point Cloud Reconstructions for Vertebral Shape Analysis. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 375-383). Springer, Cham.
[30] Drokin, I., & Ericheva, E. (2020). Deep Learning on Point Clouds for False Positive Reduction at Nodule Detection in Chest CT Scans. arXiv preprint arXiv:2005.0365
