Analyzing MRI scans to detect glioblastoma tumor using hybrid deep belief networks

Glioblastoma (GBM) is a stage IV aggressive malignant brain tumor, which is generally found in the cerebral hemispheres of the brain [1]. The treatment of GBM tumor is very difficult and cure is not possible in most of the cases. The treatment can only slow down the progress of the cancer and may reduce the symptoms and discomfort. The diagnosis of GBM contains neurological exams, imaging tests i.e. MRI and biopsy. The GBM tumor may occur at any age; however, the high majority of the patients are adults. The symptoms of GBM may include worse headache, nausea, vomiting and seizures (Fig. 1).

Medical science is equipped with advanced devices and technology. MRI machines are able to capture high contrast images of the brain and other parts of the body. These MRI scans are very useful to diagnose and detect tumors and other defected cells. However, sufficient knowledge and experience is desirable in order to read and understand these MRI scans. Sometimes, unavailability of such trained people may delay the diagnosis process. Therefore, in order to automate the process, a classification model can be developed using machine learning methods. Artificial neural network (ANN) is one of the powerful approaches in machine learning which are able to handle large amounts of data with desirable accuracy. Advancement in ANN came up with deep neural architectures i.e. convolutional neural network (CNN) and recurrent neural network (RNN). Deep neural network [2] allows us to understand the hidden patterns from the complex data and images. Deep belief networks (DBN) is a class of deep neural networks with multiple layers of hidden units in which connections are established between layers rather than units within each layer. Each layer in DBN is a restricted Boltzmann machine (RBM) stacked against each other to construct the DBN. Deep learning approaches are able to automatically extract the features from the large data sets, however the correctness of the extracted features are not guaranteed [3‐5]. In order to overcome this issue, this article proposed a DTW based DBN approach for image classification.

The rest of the article is organized as follows: In “Literature review section”, state of art literature review is presented. “Materials and methods” section discusses the materials and methods that consist of the description of the data set and methodology used in the study. “Results and discussion” section presents results and discussion and “Conclusion and future scope” section concludes the article with limitation and future scope of the article.

Deep learning approaches have gained a lot of attention in the last decade in the medical domain. Deep learning methods are capable of extracting complex features from the raw data and able to construct new features from the existing one. Affonso et al. [6] argued that human inspection is a time consuming process in several industries to check the quality of materials. Automatization could improve this process through digital image processing. However, accuracy and computational cost are very important and therefore there should be a balance between the two. The authors used convolutional neural networks (CNN) with feature descriptors to analyze real world data of woods samples from industries. They concluded that CNN along with the feature descriptor provides a better accuracy in classification of the image data. Qayyum et al. [7] proposed a CNN based approach for content-based image retrieval system for medical image classification. They used a rich data set of medical images with twenty-four classes for analysis. They proposed that learning discriminative features directly from the images may reduce the semantic gaps and would improve the classification. They achieved a classification accuracy of 99.77% with deep CNN.

Jaiswal et al. [2], used CNN to classify the pneumonia using chest X-ray. They used publically available chest x-ray dataset from RSNA (Radiological Society of North America). They evaluated several deep learning models and found that Mask-RNN is able to achieve better accuracy among others. In [8], Litjen et al. conducted a comprehensive state of art survey on various methods, applications of deep learning approaches in medical image analysis, segmentation and object detection. They illustrated how deep learning is facilitating several aspects of the medical domain. Khatami et al. [3] proposed an image classification approach for radiography images. They applied Kolmogorov–Smirnov test to find the most discriminative features from the radiography images. They argue that DTW can slightly improve the classification performance using deep networks. Another study [9] extended the deep convolutional neural network to classify the echocardiography video images to assist in the diagnosis of heart disease. Data-driven learning frameworks by incorporating both spatial and temporal information from video images have been given as input to the CNN in this study and they have achieved 92.1% accuracy. They also mentioned that incorporating feature extraction helps to improve the classification performance of deep neural networks. Gao et al. [5] proposed a deep learning based model to diagnose Alzheimer disease using CT scans of the brain. They mentioned that CNN provides desirable accuracy that helps in early diagnosis of Alzheimer disease. They developed an advanced CNN incorporating both 2D and 3D CNN networks. The stated that the modified CNN was able to achieve better accuracy and classification in comparison to other 2D CNN models. Sharma et al. [10] used a deep learning framework to diagnose cancer cells. They developed a CNN model to classify the cancer image data and compared with other traditional approaches. They also found that CNN performs better than other methods. Al-Rahhal et al. [11] developed a deep learning framework to classify the electrocardiogram signals (ECG) to detect the heart disease. The extracted suitable features from the ECG data using stacked auto-encoders. Next, a softmax regression layer is added on the top of the hidden layers in the deep neural network. Moreover, expert recommendations are considered to label the ECG signals. They mentioned that the proposed approach achieved better accuracy than other state of the art approaches.

In another study [12], Tang et al. used multi-scale representation to capture and represent features from the medical images. Further, extracted features were encoded using Fisher vector technique. The authors claimed that the proposed approach demonstrated a superior performance on the data set. Zhong et al. [13] stated that the deep learning model specifically CNN needs a sufficiently large number of labeled samples to train the model. However, real world hyper spectral labeled images are limited in the amount. Therefore, deep learning may not be suitable for such kind of data. They suggested that DBN is able to deal with this issue as it provides unsupervised pre-training to the unlabeled samples but may result in several dead hidden units. They proposed an approach to deal with this problem and obtained better performance than the original DBN. Zhao et al. [14] proposed a feature learning method known as discriminant DBN for learning high-level features for image classification. These discriminant features are learned by ensemble learning mechanism with DBN in an unsupervised manner. Wang et al. [15] suggested a semi supervised DBN based on Boltzmann machine by introducing the relevant constraint and the supervised information are equivalently integrated into the learning process of this restricted Boltzmann machine and then, this model is constructed to improve the classification accuracy. Shi and Chi [16] proposed a hyper-spectral image classification method. This method has exploited the 3D spectral-spatial information using 3D deep neural networks. They constructed a super pixel to generate the spatial image to increase the spectral-spatial similarity and diversity. Next, a 3D super pixel based sample filling method has been applied to solve the misclassification issue and finally, a 3D recurrent CNN had been used for classification of images. Similarly, Paoletti et al. [17] proposed a deep CNN based image classification model for hyper-spectral images by presenting a 3D network based CNN that uses both spatial and spectral information. They have also implemented mirroring strategy to effectively process border areas in the images. The development of computational models for medical image diagnosis raised a new level of interventional treatment. Therefore, medical image processing is now one of the most effective ways to quickly diagnose the diseases and assist the medical staff in quick decision making. As a solution, machine learning techniques are being used to automate the diagnosis process since last decade. However, traditional machine learning methods are not sufficient to deal with complex problems. Therefore, deep learning methods are trying to resolve the limitations of machine learning based techniques. In this work, an attempt has been made to show a clear cut differentiation between deep learning based methods and non-deep learning based methods for image classification.

Experiments have been performed to evaluate the performance of DWT-PCA-DBN based classifiers. All the experiments have been carried out using Python programming environment. Since, the size of the data was very huge, we have used a local cloud server with high-end configurations along with GPU support. To compare the performances among several settings, we have trained DBN with various parameters and structures and computed the results of training and testing errors for each scenario. To compare the performance with other classification techniques, various performance criterion such as classification accuracy, specificity, sensitivity and F-score are considered. The total experimentation process can be compiled in various stages as detailed below.

The following are few observations of the current study that makes it challenging and interesting for image classification.

Deep learning is well known approach for image classification as it is able to automatically extract the features from image data for further processing. However, the correctness of the extracted features is not guaranteed because there is no mathematical validation available for the correctness of those extracted features. To address this issue, this study proposed an improved DWT-PCA-DBN image classifier based image classification model by combining DWT for feature extraction and PCA to get optimized features with DBN. The DBN is composed of stacked RBMs to extract the more significant features from the reduced datasets layer by layer. Generally, DBN requires huge and multiple hidden layers with huge number of hidden units to learn the best features from the raw pixels of image data. This actually increases the complexity as well as training time for the model. Therefore, by integrating DBN with wavelet transform techniques both complexity as well as training time has been reduced. Except using raw images, the extracted low resolution images from DWT are used for training the proposed image classifier. The proposed model has been experimented with deep learning based approaches such as traditional DBN, PCA-DBN with non-deep learning based approaches of variants of ANN and also the performance of the proposed classifier has been evaluated based on number of epochs for training of RBN and weight initialization. Finally, the statistical validation justifies the efficiency of this method. The work can be extended to improve the computation efficiency of the model with respect to large amount of dataset with occlusion kind of patterns. Occlusion generally refers the blockage of blood flow into the brain. This type of pattern may occur in a tumor spot. This is also a kind of abnormality which may need attention for proper diagnosis. In this work we have not experimented our model for tumors with occlusion kind of pattern due to the unavailability of such patterns in the data. However, the work can be extended upon the availability of data with occlusion patterns using deep learning methods.