Abstract
Normal Pressure Hydrocephalus (NPH), an Atypical Parkinsonian syndrome, is a neurological syndrome that mainly affects elderly people. This syndrome shows the symptoms of Parkinson’s disease (PD), such as walking impairment, dementia, impaired bladder control, and mental impairment. The Magnetic Resonance Imaging (MRI) is the aptest modality for the detection of the abnormal build-up of cerebrospinal fluid in the brain’s cavities or ventricles, which is the major cause of NPH. This work aims to develop an automated biomarker for NPH segmentation and classification (NPH-SC) that efficiently detect hydrocephalus using a deep learning-based approach. Removal of non-cerebral tissues (skull, scalp, and dura) and noise from brain images by skull stripping, unsharp-mask based edge sharpening, segmentation by marker-based watershed algorithm, and labelling are performed to improve the accuracy of the CNN based classification system. The brain ventricles are extracted using the external and internal markers and then fed into the convolutional neural networks (CNN) for classification. This automated NPH-SC model achieved a sensitivity of 96%, a specificity of 100%, and a validation accuracy of 97%. The prediction system, with the help of a CNN classifier, is used for the calculation of test accuracy of the system and obtained promising 98% accuracy.
Keywords
Introduction
A neurodegenerative condition is a heterogeneous group of ailments distinguished by the gradual degeneration of the central nervous system’s function and structure. One of the most common neurodegenerative diseases in older people is normal pressure hydrocephalus (NPH). NPH was 20% in 70–79 years of age and 59% in 80 years and older, with no gender differences [9]. Gait disturbance, urinary incontinence, and cognitive impairment are some of the characteristics of NPH [15]. Symptoms of NPH are very similar to that of Alzheimer’s and Parkinson’s disease. The memory impairment of NPH resembles those associated with early Alzheimer’s, and the gait disturbance in NPH looks identical to those of Parkinson’s disease. People with Parkinson’s disease may have gait disorder, incontinence, and dementia associated with NPH, but they hardly have the enlarged ventricles [10]. Hydrocephalus is a condition by which the brain ventricles become enlarged because of the accumulation of cerebrospinal fluid (CSF) within the ventricles or cavities. Accumulation occurs when CSF’s normal flow is blocked, which leads to excess CSF and results in Hydrocephalus.
Imaging modalities play a significant role in evaluating patients with anatomical disorders. These non-invasive studies test blood flow from tissues, brain metabolism, water movement, and blood oxygen level-dependent with biomarkers’ aid [2]. Popular imaging modalities like magnetic resonance imaging (MRI), ultrasound, and computed tomography (CT) can identify brain ventricles’ enlargement. The structural MRI (sMRI) detects brain structural changes because of the enlarged ventricles and CSF flow in NPH. MRI uses radio waves and a pure magnetic field to capture and examine detailed images of the brain’s organs, tissues, anatomy, and pathology.
MRI images significantly impact automated medical analysis to maintain much information about the abnormalities in the brain tissues and brain structures [12]. The non-invasive technique, structural magnetic resonance imaging (sMRI), can examine the brain’s anatomy and pathology. Various aspects of brain morphology and function are obtained by scanning using sMRI technology. This method, along with the biomarkers, can identify the neurodegeneration and changes in microstructure. The computer analyzes beneficial representations and capabilities from the raw data in deep learning. The main characteristic of deep learning is feature learning, which learns the data representations [11]. This paper’s contribution is applying the deep learning concept to develop an automated NPH-SC model classification using brain sMRI images. The proposed method aims to differentiate between normal brain and NPH using structural brain MRI images with better performance. Some popular techniques used for detecting NPH and similar diseases using image processing techniques are shown in Table 1.
Overview of recent studies that have used different techniques for the diagnosis of NPH
Overview of recent studies that have used different techniques for the diagnosis of NPH
A parametric model for the differential diagnosis of normal pressure hydrocephalus distinguishes NPH from Alzheimer’s disease (AD), PD, and dementia with Lewy bodies (DLB) [8]. Using the mean diffused (MD) shape parameters, they distinguished NPH from 3 other disorders with 96% specificity and 86% sensitivity. When differentiated from NPH and AD, the same technique achieved 88% specificity and 86% sensitivity [4]. The combination of mathematical and statistical pre-processing methods, morphological operation, and a priori knowledge accurately classifies white matter, subarachnoid space, and brain ventricles by a random forest classifier, with no gender differences. An automated classification model to classify scans for tightening sulci in the brain’s high convexities obtained an area under the receiver operating characteristic (AUROC) of 0.99 using a support vector machine with selected sulcal CSF volumes as to its features [9]. A fully automated model with a 3D deep learning classifier differentiates NPH, AD, and control subjects. They evaluated the performance tested using a leave-one-out cross-validation test and achieved an accuracy of 90%. They also assessed the result’s validity, using the gradient-weighted class activation mapping (Grad-CAM) technique of visualizing the essential areas in distinguishing NPH and AD on the input image [10]. A method was developed to segment and label ventricles in NPH patients from MRI images by combining a patch-based tissue classification with a registration-based multi-alias labelling method, which has provided a robust segmentation and labelling in terms of overlap measure with dice coefficient in it [8]. A modified 3D U-net to perform ventricular parcellation by MRIs has achieved mean dice similarity coefficient (DSC) of 0.895±0.03 for the ventricular system and a mean DSC of 0.973±0.02 on the NPH data sets was obtained [11].
The present system lacks an exact diagnosis and classification due to the similarities and common symptoms of AD and NPH. The primary aim of the proposed work is to develop an automated biomarker for NPH diagnosis. The proposed automated method enables us to achieve better accuracy for classification, reduces the diagnostic time of radiologists, and increases diagnosis accuracy.
Imaging modalities like MRI, computed tomography(CT), and physical examination diagnoses the NPH syndrome. Patients who show clinical improvement after lumbar puncture will be candidates for endoscopic third ventriculostomy (ETV) or shunt surgery. Although lumbar puncture has less negative prognosis value (less than 20%), patients improve following shunt surgery even with a specific negative lumbar puncture test. Therefore, a more useful diagnostic method for identifying NPH patients is required. This proposed work comprises image acquisition, pre-processing sMRI images, segmentation of ventricles, and classification of brain sMRI as NPH or normal using CNN architecture.
Proposed method
In the axial T1-weighted sMRI, the ventricle’s structural deformations are visible, so we collected these images from Anugraha Neurocare-Trivandrum, India. Fifty normal and fifty NPH cases are collected for the study. The proposed methodology, automated diagnosis of NPH, consists of modules such as image acquisition, skull stripping, edge enhancement, segmentation, and classification. Figure 1 shows the block diagram of the proposed methodology.
Block diagram of the Proposed Methodology.
After image acquisition, image pre-processing is done to remove the unwanted labels and noises present in the images. The acquired sMRI consists of the brain and non-brain regions, so in the next step, the non-brain regions are removed using skull-stripping [17]. Then the unsharp mask filter is implemented to enhance the edges in different regions of sMRI, and thereby the segmentation becomes more easy and accurate.
The major steps used in the skull stripping module include; Otsu’s binarization, mask creation, and superimposition. The sMRI is converted to a binary image using binarization. The otsu method automatically performs depletion of a gray level image to a binary image. The image is considered as two classes of pixels, background, and foreground. The optimal threshold for classifying the two categories is obtained so that their inter-class variance is maximal and intra-class variance is minimal. The second step in this module is creating the mask to extract the largest connected component, the brain, from the binary image. The dilation and erosion operations are performed to extract the mask, and these operations preserve the minute characteristics of the brain in the resultant image. The created mask is superimposed on the brain images to extract the skull stripped image.
Skull stripped images are then pre-processed by Unsharp mask filtering to enable more prominence to the edges. This filter improves the edges and other frequency components in images by deducting a smoothed or unsharp version of the images from the input images. Here, in this study, an unsharp mask filter reduces the noise and sharpens the edges. The number of neighboring pixels on the affected edges is chosen by setting a “Radius” parameter in the filtering process. The darkness of the edges is specified appropriately by a “percentage” parameter. Distance between neighboring tone values is done by developing a “Threshold” parameter before the filter does anything. Thus the pre-processed images are obtained after performing skull stripping and unsharp mask filter.
Watershed segmentation
The watershed segmentation method is used in this work to extract the ventricles. This method is applied to segment the ventricle borders. Here, the otsu binarized images are taken for segmentation. From the otsu binarization, morphological opening and closing operations are done. Initially, the far-away regions from the objects are considered background and the nearby objects as foreground. The regions which are neither foreground nor background are considered as the boundary. The boundary pixels removes unwanted areas by erosion. Dilation finds the areas that are not sure and excludes them. In this way, we could identify the brain ventricle in the foreground region. A Watershed marker is created (An array of int32 data types, with the same original image size) and the regions inside are labeled. The known regions (whether foreground or background) are marked with one of the positive integers, and the unknown areas are left as zero. Thus, the marker is created around the ventricles in the foreground region.
Convolutional neural networks
Classifying an image according to its visual content, is referred to as image classification. The process of classification is to sort all pixels in a digital image into several classes. A convolutional neural network (CNN) is used for the classification of segmented images. There are dimensions in CNN in which the layers are organized into; width, depth, and height. The neurons in one layer are not linked to all the neurons in the next layer, but only to a small portion. The final output is then limited to a single vector of a score with the depth dimension. The proposed convolutional network consists of four convolutional layers with kernel size 3*3 and ReLu activation function, followed by max-pooling, dense, dropout, and fully connected layers. The two major parts of CNN include; feature extraction and classification. In feature extraction, a series of convolutions and pooling operations are performed by the network by which the features could be detected. The fully connected layers act as classifiers on the extracted components.
In convolution, one function changes the other by integrating two functions. The feature map, feature detector, and the input image are the three key processes involved in this classification. The image detected will be the input image. In this work, the feature detector is a matrix of 3*3, otherwise known as a filter or kernel. The feature map is obtained by the matrix representation of the input image, multiplying element-wise with the previous feature detector known as an activation map or convolved features. The primary purpose of this step is to reduce the image size and make processing easier and faster.
Here the padding is set to 1 as it is the number of pixels added to an image. We also set stride to 1 as it is the pixel shifts over the input matrix. To increase the non-linearity in CNN images, we implement the activation function ReLu (rectified linear unit).
To detect the features from the images pooling operation is done in different angles of images and differences in pictures. A matrix of 2*2 is placed on the feature map that picks the largest value in that box. Max-pooling is performed by placing a matrix of 2*2 on the feature map. The above matrix 2*2 moves from left to right throughout the feature map and picks the largest value in each pass. Those values form a new matrix called pooled feature maps. We perform max-pooling to reduce the image size and preserve the key features of the image. This technique can minimize overfitting.
We transform the entire pooled feature map matrix into a single column by flattening it and then fed it into the network for further processing. After Flattening, the flattened feature map passes through the neural network. This level is composed of an input layer, an output layer, and a fully connected layer and is the stage where the process of classification begins. The layer at which it predicts classes is the output layer. Images are pre-processed before passing them to a predicted method, which prevents overfitting. We add the fully connected layer to the neural network by dense. This layer connects all the neurons to the next layer dropout, preventing the model from overfitting. The fully connected layer classifies the input image into various classes based on the training set. An accuracy of 97% is achieved after the classification of NPH and Normal sMRI images.
Results
The proposed method aims to provide accurate segmentation of the ventricles and the classification of NPH and non-NPH patients. Hundred patient cases were selected for the studies among fifty were normal, and fifty were NPH affected. The axial view of sMRI images was taken for the study because the ventricular regions were clearly visible in the axial view.
Enlarged brain ventricle by the accumulation of cerebrospinal fluid is the primary cause of NPH. Patients with age group 70 to 85 were taken for this study with no gender difference. In this study, we took sMRI slices for a patient with various slice thicknesses. Figure 3 shows different slices of axial views of sMRI of Normal and NPH affected patients.
CNN Architecture.
sMRI of Normal and NPH affected patients.
In this study, the skull stripping-preprocessing technique was carried out to remove the unwanted areas in the sMRI brain images. Figure 4(b) shows the results after performing skull stripping in the input image Fig. 4(a). The image in Fig. 4(c) shows the results of the Unsharp mask filter. By applying this filter, the edges of images get sharpened, providing fine details of the image.
Sample results after segmentation and classification (a) Input Image (b) Skull-stripped image (c) Filtered Image (d) Segmentation by Watershed transform.
Brain ventricles were segmented by marker-based watershed segmentation and are shown in Fig. 4(d). Performance evaluation of watershed segmentation was calculated using the Jaccard Index/Intersection over Union to understand the similarities between the two data sets. An accuracy of 93% is achieved in segmentation by performing the Jaccard Index.
In CNN architecture, the feature maps were the outputs obtained representing the intermediate layers after the 1st layer. Figure 5 shows the output obtained by the model that has been used for validation and training.
Feature maps for sMRI brain image.
Visualization of the CNN model helps us to understand the model, thereby adjusting the model’s parameters for better classification. We reported the classification evaluation on the 50-image test set. The ideal confusion matrix (CM) was
The classification results show that the 1st graph (Fig. 6(a)) is the accuracy of training and validation datasets over training epochs. We perform 50 epochs. If we try training for more epochs, overfitting occurs, which means the model recognizes the training set’s specific images. The 2nd graph (Fig. 6(b)) is the plot of loss of training and validation datasets. Here, the validation error is low but slightly higher than the training error. It means the network is a good fit. An accuracy of 97% is achieved after classification. The loss function, binary cross-entropy, is used in binary classification tasks to predict test images. Fifty test images were used for prediction.
Accuracy and Loss of Training and Validation.
The Precision-Recall method is one of the metrics used for the evaluation of the binary classification model. For each possible threshold, the relationship between precision (positive predictive value) and recall (sensitivity) is shown in Fig. 7. The x-axis shows the recall, and the y-axis shows the precision for varying thresholds. This curve describes how good a model is at predicting the positive classes. Table 2 shows the comparison of different methods that have been used for the detection of NPH syndrome.
Precision-Recall curve.
Comparison of different methods for the detection of NPH syndrome
In this paper, we successfully classified the NPH data from normal control with 97% accuracy using CNN deep learning architecture, which was trained and tested with a set of hundred patient cases. Experimental data was obtained from Anugraha Neurocare, Trivandrum, India. T1-weighted axial sMRI images were first preprocessed using skull-stripping and unsharp mask filter to get more clear, label-free and skull stripped image. The skull stripped images were segmented again using the watershed segmentation technique to extract the brain ventricles. The study extracted the most appropriate features from the segmented ventricle regions and done classification using CNN. This deep learning solution enables researchers and physicians to predict any new data potentially. The accuracy achieved in this work was very high, confirming the network architecture was correctly selected. However, more complicated network architecture, including more convolutional neural layers or new architecture, is recommended for future works. More clinical data in different views can also be incorporated in future work to build a robust system.