(PDF) Deep AutoEncoder-Decoder Framework for Semantic Segmentation of Brain Tumor

ICONIP2019 Proceedings 53 Deep AutoEncoder-Decoder Framework for Semantic Segmentation of Brain Tumor Arshia Rehman1 Saeeda NazB1 , Usman Naseem2 , Imran Razzak2 , and Ibrahim A Hameed3 1 Computer Science Department, GGPGC No.1, Abbottabad, Pakistan 2 University of Technology, Sydney, Australia 3 Norwegian University of Science and Technology

B Abstract. Accurate segmentation of brain tumor is a critical component for diagnosis of cancer, treat- ment and evaluation of outcome. It consist of identification of different types of tumor tissues from normal brain MRI images. Recently, pathway CNNs have been used for semantic segmentation, however are com- putationally expensive. Build upon success of SegNet, in this paper, we presented different architectures of SegNeT encoder and decoder based on pixel-wise classification. Nonlinear up sampling are performed by the model. The end to end training and small number of parameters used for the training makes the computational process more higher than other deep learning architectures. We performed the semantic segmentation on the MRI brain tumor Figshare - dataset and achieved the state of the arts results (99.93% global accuracy) in comparison to traditional CNN models. 1 Introduction One of the core and complex organ of human body is brain that comprises nerve cells and tissues to control the foremost activities of the entire body like breathing, movement of muscles and our senses. Every cells have their own capabilities, some cell grows with their own functionality and some lose their capability, resist and grow aberrant. These mass collections of the cells form the tissue is known as “tumor”. Cancerous brain tumors are uncontrolled and unnatural growth of brain cells that cause the damage of nervous system severely results painful death of patient. Although, it is not very common disease, however, is one of the most life threatening and lethal cancers. For example, in 2015, approximately 23,000 patients were diagnosed brain tumor only in United States. According to 2017 cancer statistics [15], brain tumor is measured as one of the foremost cause of cancer-related indisposition, morbidity, and mortality around the world both in children and in adults [18, 24]. The goal of brain tumor segmentation is to identify the brain tumor and extract the patient specifically clinical information to help later interventions that exist in multidimensional Magnetic Resonance Imaging (MRI) images. Gliomas are infiltrative in nature and one of the most common tissues in brain. They are difficult to identify and segment as can spread to any part of brain. High-grade gliomas is the one of the aggressive brain tumors with a median survival of almost two years. Brian tumor surgery is one of the commonly applied treatment, however, chemotherapy and radiation can also be used to reduce the growth of tumors that are complex to remove through surgery [20]. Thus, accurate identification and segmentation of brain tumor has great impact on improving the treatment and its planning i.e. detection and segmentation of tumor cells not only help to identify the existence of cancerous cells only but also provide valuable information of their location, shape, size, and difference between tissues i.e. necrotic tissue, edema (swelling near the tumor) and tumorous tissue (vascularized or not). Although some of the brain tumors (i.e. meningiomas) are easy to segment, however, there are some that are difficult to identity (i.e. gliomas) due to their complex properties i.e. tumor edges are often ambiguous and can not be differentiated from healthy tissues due to their fuzzy nature [17]. Thus, there is a need to sensitive and accurate methods to identify the tumors in order to increase the survival rate. Manual segmentation from high dimensional large MRI data is challenging, time-consuming, tedious, prone to error and affected by inter-observer variability [19]. Hence, physicians often use qualitative or visual inspection of tumor only, or maximum they use the crude measures like approximating the tumor volume and their quantity. Automated analysis and quantitative assessment of tumor cells provides valuable information for the early diagnosis and help to plan the early treatment strategies. Existing approaches can be classified into generative models or discriminative models. Generative model identify the cancerous cell by differentiating healthy cells from tumors based on their appearance. Thus, requires knowledge of the anatomical structure of brain that can be computed by aligning healthy tissues and affected image i.e. contour-based segmentation requires alignment- based features or left-right brain symmetry features in order to align the image. It is one of the affecting technique and does not require label data to train. Unlike generative models, the discriminative segmentation methods require extensive size of label data as they use prior knowledge of brain i.e. segmentation based on low-level image features such as raw pixels and Ga- bor filter banks [22, 23]. Traditional discriminative methods for brain tumor segmentation include conventional Volume 15, No. 4 Australian Journal of Intelligent Information Processing Systems 54 ICONIP2019 Proceedings machine learning approaches such as neural networks and support vector machines etc [16]. that requires hand- crafted features consisting of high discriminative power. Automatic features learned by deep neural networks outperformed as compare to handcrafted features [21, 23, 6]. The model based features are more abstractive and discriminative extracted automatically through the deep learning models. Automated cancer detection from MRI images is a well studied task in the computer vision community; however, little work exists in differentiating the segmented tumor region using pixel-wise semantic segmenation. Pre-invasive segmentation presents a more difficult classification scenario than the binary classification task of invasive cancer detection. It necessitates careful analysis of epithelial and fuzzy structures in the MRI brain images. This paper presents the semantic segmentation using well-known model SegNet to produce a tissue label image for the MRI brain images that can lead to an automated diagnostic system. Features maps are generated by the encoder depth-4 and VGG in the SegNet architecture. Evaluation is on two classes tumor and non-tumor for the segmentation. Motivation of the work is to gain the more efficient performance for the brain tumor detection with less memory usage and the improved computational time. The proposed approach is the best fit for the segmentation of the objects and resulted into the high performance rate. The SegNet architecture is explained in detail, and set of experiments are conducted to compare its segmentation performance to other models. Rest of the paper is organized as, Section II provides the related work, followed by the section III about architecture of the proposed approach. In section IV, comprehensive evaluation on brain tumor dataset - Figshare is explained in the light of experiments conducted. Finally Section V draws the conclusion with future direction. 2 Related Work Initial work of semantic segmentation have focused on designing a robust feature representation, e.g. TextonFor- est [26] TextonBoost [27], as well as Random Forest-based classifiers [25]. Recently, deep convolutional learning models have observed extensive success in many domains and successfully employed for semantic segmentation. In particular, a number of CNN architectures like DeepLab [7], RefineNet [11], Fully Convolutional Networks (FCN) [12] and SegNet [5] have shown significant progress in performance and accuracy by adapting the deep Convolutional Neural Networks (CNN) based image classifiers to semantic segmentation [14]. Among the vari- ous CNN models, SegNet has been employed on outdoor road scene and indoor scene images. Mehta et al. [13] enhanced the SegNet architecture for the segmentation of breast biopsy whole slide images. The SegNet archi- tecture also yielded best accuracy for segmentation of blood cells (white blood and red blood cells) and the background in the blood smear images using ALL-IDB1 dataset in [29]. Inspired from these best performance achieved in [5, 13, 29], we are going to deploy the SegNet architecture on MRI brain images for the segmentation of brain tumor. The proposed system relay on the MRI images of Brain Tumor Dataset - Figshare thus, we here present the related work of figshare dataset. The used techniques on the same dataset is the key to the equitable comparison between the different methods. Rehman et al[22] performed deep transfer learning techniques using three CNN models: AlexNet, GoogLeNet, and VGGNet. They conducted experiment on augmented dataset and achieved 98.69% accuracy on VGG16. Cheng et al. [8] first conducted experiment on Brain Tumor Dataset - Figshare. They used augmented tumor region as region of interest and split these regions into sub regions by employing adaptive spatial division method. Authors have extracted the intensity histogram, bag-of-words (BoW), gray level co-occurrence matrix (GLCM) based features. They reported highest accuracy of 87.54%, 89.72%, and 91.28% on extracted features using ring-form partition method Another contribution of same work was presented in [9]. They deployed Fisher Vector for the aggregation of local features of each sub region. Mean average precision (map) 94.68% was retrieved. Ismael and Abdel-Qader [10] extracted statistical features from MRI images of Brain Tumor Dataset - Figshare with the aid of 2D Discrete Wavelet Transform (DWT) and Gabor filter techniques. They classified using back propagation multi-layer perceptron neural network and retrieved highest accuracy of 91.9%. Abir et al. [2] deployed Probabilistic Neural Network for classification of brain tumors. They performed image filtering, sharpening, resize and contrast enhancement in pre-processing and extracted GLCM features. They attained highest accuracy of 83.33%. Abiwinanda et al. [3] identified three common types of brain tumor. They employed five different architectures of CNN and reported highest accuracy on architecture 2. The architecture 2 comprises of 2 convolutional layers, Relu layer and max-pool followed by 64 hidden neurons. They achieved 98.51% and 84.19% on training and validation sets respectively. Afshar et al. [4] proposed a novel model Capsule networks (CapsNets) for the detection of brain tumor. They varies the feature maps in the convolutional layer of CapsNet in order to increase accuracy. They achieved highest accuracy of 86.56% using 64 feature maps with one convolutional layer of CapsNet. Widhiarso et al. [30] computed GLCM and fed to the Convolutional Neural Network. They claimed that GLCM combined with contrast feature gave 20% improved accuracy. They achieved highest accuracy of 82% using this scenario. Apart from the conventional and deep learning classification techniques, Sobhaninia et al.[28] first used the Brain Tumor Dataset - Figshare for semantic segmentation task. They employed LinkNet Australian Journal of Intelligent Information Processing Systems Volume 15, No. 4 ICONIP2019 Proceedings 55 for the segmentation of brain tumors. They deliberate MRI images of figshare dataset from different angles and deployed multiple models for segmentation. They reported 0.73 dice score using single network and 0.79 dice score using multiple networks respectively. 3 Brain Tumor Segmentation and Classification based System In this section, we present deep convolutional encoder-decoder architectures for semantic segmentation of brain tumor in detail. The semantic segmentation is the procedure of assigning each pixel of image to appropriate target label using region of interest image as a ground truth. In this study we consider SegNet as a based method and presented different architectures of SegNeT encoder and decoder based on pixel-wise segmentation of brain tumor and non-tumor region. The work flow of proposed system is depicted in Figure 1. The brain tumor T1-weighted MRI images of 233 patients (Brian Tumor Figshare - dataset) are used. MRI images are enhanced using contrast stretching technique. In the next step, SegNet encoder architectures are employed to extract the visual features from MRI images. Then last fully connected layer of encoder is removed and the resultant high dimensional feature produces the semantics segmentation mask. Finally, decoder is connected to a softmax classifier which classifies each pixel. These steps are elaborated in the next sections. 3.1 Encoder The images of size 512 × 512 are fed to the encoder, which produces feature maps through convolution with filter bank. Encoder process is same like the architecture of VGG16 that consist of 13 convolutional layers (7 × 7 ), batch normalization layer, element-wise RELU, drop out and max pooling (2 × 2), non-overlapping stride by 2, and up-sampling (2 × 2 ). Max-pooling is used to achieve translation in-variance over small spatial shifts in the image, combine that with sub-sampling. It leads to each pixel governing a larger input image context (spatial window). Although, more number of max-pooling layer and sub-sampling can provide better translation that could improve classification, however, this results in the loss of feature maps spatial resolution. Thus, before the sub-sampling, it is necessary to capture the boundary information. For this purpose, we have captured max-pooling indices only. 3.2 Decoder For each of the 13 encoders, there is a corresponding decoder which up-samples the feature map using mem- orized max-pooling indices. It takes the feature map from the encoder network as input and up-samples its corresponding input feature maps using the memorized max-pooling indices from the encoder feature maps. The generated feature map is sparse and high in resolution so that best fit for the genuine input. Dense feature maps is obtained through convolution with a trainable decoder filter bank. The resultant high dimensional features produce the semantics segmentation mask. 3.3 Softmax Classifier The last decoder is connected to a softmax classifier which classifies each pixel. The convolutional layer of 1 × 1 with the softmax evaluates the probabilities for classes. The soft-max generates the output with the probabilities of channel k where the numbers of classes are identified as k. The maximum probability of pixel wise segmentation is the result of the predicted segmentation. The correctly pixel wise segmentation for the tumor and non-tumor classes results are presented in the results section. 4 Experimental Analysis In this section, we analyze the evaluation of proposed brain tumor semantic pixels based segmentation and classification using SegNet architectures. We conducted two studies to evaluate the performance of different architectures for encoder step of SegNet model using Endcoder-4 architecture and VGG architecture of SegNet. We conducted various experiments using different parameters for getting optimal set of parameters, on which network earn optimally. Finally, the following two studies conducted using optimal parameters. 4.1 Brain Tumor Dataset - Figshare In this work, we used publicly available brain tumor ”Figshare” dataset [1] in order to analyze and evaluate the performance of proposed approach. In order to improve the performance, in this work we have done two different updates in the architectures of SegNet. The dataset comprises 3064 brain MRI slices collected from Volume 15, No. 4 Australian Journal of Intelligent Information Processing Systems 56 ICONIP2019 Proceedings Fig. 1. Pipe-line of Brain Tumor segmentation and Classification based on SegNet [5] . Australian Journal of Intelligent Information Processing Systems Volume 15, No. 4 ICONIP2019 Proceedings 57 Fig. 2. Progress Graph of Training and Validation using Brain Tumor Dataset - Figshare 233 patients with three kinds of brain tumors: Meningioma, Glioma, and Pituitary. The dataset consist of 708, 1426 and 930 number of images/slices for Meningioma, Glioma and Pituitary tumor. The dataset is publicly available on Figshare website. The dataset consist of unique labels that demonstrates the type of brain tumor, 512 ×512 image data in uint16 format, vector containing tumor border with the coordinates of discrete points, and ground truth in binary mask image. In our experiments, CNN model takes image along with its ground truth as input unit thus we extract the image data and tumor mask image data from the .mat files. For the estimation of the best model training set is used and the validation for the error prediction on the model and for the generalizability of errors test set is used. Splitting for the dataset for the training, testing and validation is as shown in the Table 1. Table 1. Statistics of Brain Tumor Dataset - Figshare Sets Percentage No of Images Training 70% 2146 Validation 15% 259 Testing 15% 259 4.2 Results and Discussion In first study, encoder depth of 4 is evaluated. The depth of network determines the number of times the input image is down-sampled or up-sampled as the input data is processed further. The encoder network down-samples the input image by a factor of 24 . The decoder network up-samples the encoder network output by a factor of 24 . In second study, pre trained architecture of VGG16 is implemented in the encoder of SegNet. We have trained the network with sgdm solver. The highest global accuracy of 99.93% when trained using ’sgdm’ solver with the batch size of 1, initial learn rate of 0.15, validation frequency of 500. The confusion metrics of each experiment are shown in Figure 3. We report different measurement metrics like global, mean, mean IOU, Weighted IOU, meanBFScore and Dice in table. 2 for semantic segmentation and classification. We have analyzed and compared the performance of proposed system with existing state of the art Brain Tumor’s detection systems on Brain Tumor Dataset - Figshare as explained in Sect. 2. A meaningful comparison of our system is possible with work of Sobhaninia et al. [28]. They first used the Brain Tumor Dataset - Figshare for semantic segmentation task. They employed LinkNet for the segmentation of brain tumors. They Volume 15, No. 4 Australian Journal of Intelligent Information Processing Systems 58 ICONIP2019 Proceedings Table 2. Result Evaluations on vgg16 and encoder depth 4 Model Overall Accuracy Network Global Mean MeanIOU Weighted IOU MeanBFScore Dice Encoder Depth 4 99.92 93.61 76.13 97.14 91 90.8 VGG16 99.93 93.65 76.22 97.19 91.04 93.14 Fig. 3. The confusion matrix of each class using encoder depth 4 and VGG deliberate MRI images of Brain Tumor Dataset - Figshare from different angles and deployed multiple models for segmentation. They reported 0.73 dice score using single network and 0.79 dice score using multiple networks respectively. Our work is the pioneer study to explore the SegNet architecture using three studies: encoder- depth-4 and VGG16. We have investigated the number of parameters to select the best network parameters for the model which resulted into the low error rate. We have attained 93.08 dice score and global accuracy of 99.92% using SegNet-encoder-depth-4, 93.14% dice score and global accuracy of 99.93% using SegNet-VGG16. Table 3. Comparison with Existing Deep Learning based Systems Reference Model Performance Measurement Sobhaninia et LinkNet 0.73 dice score using single al.[28] network and 0.79 dice score using multiple networks Our Proposed SegNet 0.9340 accuracy or 0.9314 dice score 5 Conclusion We presented the semantics segmentation architecture SegNet for the pixel-wise label segments. The SegNet produces good results for the segmentation and more accurate for the label predictions. It provided efficient performance with promising accuracy as compared to the patch based classification. It produces high results without the use of the post-processing CRF which elaborated the model to more time consuming without achieving the desired results. SegNet improves the computational time and memory as compared to the other segmentation models. The key point in the deployment of SegNet is the use of the less number of parameters. The high number of parameters for the model may cause interruptions in the network. The less number of parameters leads to the improved network efficiently. SegNet enhanced both the quality and quantitative analysis for the brain tumor segmentation. Our presented SegNet based system gives the state of the art results as compared to the other studies using brain tumor dataset - Figshare and achieved 99.93% accuracy. The experiments on the Brain Tumor Dataset - Figshare indicates the quality results over the medical imaging. In future, SegNet architecture will be explored for the brain tumor using different combinations of layers. SegNet model will also be tested on another benchmark dataset for the brain tumor segmentation and classification. We also aim to explore FCN and Unet for this problem in future. References 1. Cheng, Jun brain tumor dataset. figshare. dataset. https://doi.org/10.6084/m9.figshare.1512427.v5, accessed: 2018- 05-30 Australian Journal of Intelligent Information Processing Systems Volume 15, No. 4 ICONIP2019 Proceedings 59 2. Abir, T.A., Siraji, J.A., Ahmed, E., Khulna, B.: Analysis of a novel mri based brain tumour classification using probabilistic neural network (pnn) (2018) 3. Abiwinanda, N., Hanif, M., Hesaputra, S.T., Handayani, A., Mengko, T.R.: Brain tumor classification using con- volutional neural network. In: World Congress on Medical Physics and Biomedical Engineering 2018. pp. 183–189. Springer (2019) 4. Afshar, P., Mohammadi, A., Plataniotis, K.N.: Brain tumor type classification via capsule networks. arXiv preprint arXiv:1802.10200 (2018) 5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence 39(12), 2481–2495 (2017) 6. Bibi, Kiran Naz, S., Rehman, A.: Biometric signature authentication using machine learning techniques: Current trends, challenges and opportunities. Multimedia Tools and Applications (2019) 7. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence 40(4), 834–848 (2018) 8. Cheng, J., Huang, W., Cao, S., Yang, R., Yang, W., Yun, Z., Wang, Z., Feng, Q.: Enhanced performance of brain tumor classification via tumor region augmentation and partition. PloS one 10(10), e0140381 (2015) 9. Cheng, J., Yang, W., Huang, M., Huang, W., Jiang, J., Zhou, Y., Yang, R., Zhao, J., Feng, Y., Feng, Q., et al.: Retrieval of brain tumors by adaptive spatial pooling and fisher vector representation. PloS one 11(6), e0157112 (2016) 10. Ismael, M.R., Abdel-Qader, I.: Brain tumor classification via statistical features and back-propagation neural net- work. In: 2018 IEEE International Conference on Electro/Information Technology (EIT). pp. 0252–0257. IEEE (2018) 11. Lin, G., Milan, A., Shen, C., Reid, I.: Refinenet: Multi-path refinement networks for high-resolution semantic seg- mentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1925–1934 (2017) 12. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431–3440 (2015) 13. Mehta, S., Mercan, E., Bartlett, J., Weaver, D., Elmore, J., Shapiro, L.: Learning to segment breast biopsy whole slide images. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 663–672. IEEE (2018) 14. Naseer, A., Rani, M., Naz, S., Razzak, M.I., Imran, M., Xu, G.: Refining parkinson’s neurological disorder identifi- cation through deep transfer learning. Neural Computing and Applications pp. 1–16 (2019) 15. R, S., C R, M., A., J.: Cancer statistics, 2017. CA: A Cancer Journal for Clinicians 67(1), 7–30 (2017) 16. Razzak, I., Blumenstein, M., Xu, G.: Multiclass support matrix machines by maximizing the inter-class margin for single trial eeg classification. IEEE Transactions on Neural Systems and Rehabilitation Engineering (2019) 17. Razzak, I., Imran, M., Xu, G.: Efficient brain tumor segmentation with multiscale two-pathway-group conventional neural networks. IEEE journal of biomedical and health informatics (2018) 18. Razzak, M.I., Imran, M., Xu, G.: Big data analytics for preventive medicine. Neural Computing and Applications pp. 1–35 (2019) 19. Razzak, M.I., Naz, S.: Microscopic blood smear segmentation and classification using deep contour aware cnn and extreme machine learning. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 801–807. IEEE (2017) 20. Razzak, M.I., Naz, S., Zaib, A.: Deep learning for medical image processing: Overview, challenges and the future. In: Classification in BioApps, pp. 323–350. Springer (2018) 21. Rehman, A., Naz, S., Razzak, M.I.: Writer identification using machine learning approaches: a comprehensive review. Multimedia Tools and Applications 78(8), 10889–10931 (2019) 22. Rehman, A., Naz, S., Razzak, M.I., Akram, F., Imran, M.: A deep learning-based framework for automatic brain tumors classification using transfer learning. Circuits, Systems, and Signal Processing pp. 1–19 (2019) 23. Rehman, A., Naz, S., Razzak, M.I., Hameed, I.A.: Automatic visual features for writer identification: a deep learning approach. IEEE Access 7, 17149–17157 (2019) 24. Rehman, A., Naz, S., Razzak, M.I., Hameed, I.A.: Leveraging big data analytics in healthcare enhancement: Chal- lenges and opportunities. Future Generation Computer Systems (2020) 25. Shotton, J., Fitzgibbon, A., Cook, M., Sharp, T., Finocchio, M., Moore, R., Kipman, A., Blake, A.: Real-time human pose recognition in parts from single depth images (2011) 26. Shotton, J., Johnson, M., Cipolla, R.: Semantic texton forests for image categorization and segmentation. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1–8. IEEE (2008) 27. Shotton, J., Winn, J., Rother, C., Criminisi, A.: Textonboost: Joint appearance, shape and context modeling for multi-class object recognition and segmentation. In: European conference on computer vision. pp. 1–15. Springer (2006) 28. Sobhaninia, Z., Rezaei, S., Noroozi, A., Ahmadi, M., Zarrabi, H., Karimi, N., Emami, A., Samavi, S.: Brain tumor segmentation using deep learning by type specific sorting of images. arXiv preprint arXiv:1809.07786 (2018) 29. Tran, T., Kwon, O.H., Kwon, K.R., Lee, S.H., Kang, K.W.: Blood cell images segmentation using deep learning semantic segmentation. In: IEEE International Conference on Electronics and Communication Engineering-(ICECE). pp. 13–16 (2018) 30. Widhiarso, W., Yohannes, Y., Prakarsah, C.: Brain tumor classification using gray level co-occurrence matrix and convolutional neural network. IJEIS (Indonesian Journal of Electronics and Instrumentation Systems) 8(2), 179–190 (2018) Volume 15, No. 4 Australian Journal of Intelligent Information Processing Systems 60 ICONIP2019 Proceedings 23. Zafar Saeed, Rabeeh Ayaz Abbasi, Imran Razzak, Onaiza Maqbool, Abida Sadaf, and Guandong Xu. En- hanced heartbeat graph for emerging event detection on twitter using time series networks. Expert Systems with Applications, 2019. 24. Zafar Saeed, Rabeeh Ayaz Abbasi, Muhammad Imran Razzak, and Guandong Xu. Event detection in twitter stream using weighted dynamic heartbeat graph approach. arXiv preprint arXiv:1902.08522, 2019. 25. Zafar Saeed, Rabeeh Ayaz Abbasi, Abida Sadaf, Muhammad Imran Razzak, and Guandong Xu. Text stream to temporal network-a dynamic heartbeat graph to detect emerging events on twitter. In Pacific-Asia Conference on Knowledge Discovery and Data Mining, pages 534–545. Springer, 2018. 26. M. Schuster and K.K. Paliwal. Bidirectional recurrent neural networks. Trans. Sig. Proc., 45(11):2673–2681, November 1997. 27. Sara Owsley Sood, Judd Antin, and Elizabeth F. Churchill. Using crowdsourcing to improve profanity detection. In Wisdom of the Crowd - Papers from the AAAI Spring Symposium, AAAI Spring Symposium - Technical Report, pages 69–74, 8 2012. 28. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15:1929–1958, 2014. 29. Zeerak Waseem and Dirk Hovy. Hateful symbols or hateful people? predictive features for hate speech detection on twitter. In SRW@HLT-NAACL, 2016. 30. Ellery Wulczyn, Nithum Thain, and Lucas Dixon. Ex machina: Personal attacks seen at scale. In Proceedings of the 26th International Conference on World Wide Web, pages 1391–1399. International World Wide Web Conferences Steering Committee, 2017. 31. Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alexander J. Smola, and Eduard H. Hovy. Hierarchical attention networks for document classification. In HLT-NAACL, 2016. 32. Dawei Yin, Zhenzhen Xue, Liangjie Hong, Brian D Davison, April Kontostathis, and Lynne Edwards. Detection of harassment on web 2.0. Proceedings of the Content Analysis in the WEB, 2:1–7, 2009. Australian Journal of Intelligent Information Processing Systems Volume 15, No. 4