Hybrid improved fuzzy C-means and watershed segmentation to classify Alzheimer’s using deep learning

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 13, No. 4, December 2024, pp. 4080~4094
ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i4.pp4080-4094  4080
Journal homepage: http://ijai.iaescore.com
Hybrid improved fuzzy C-means and watershed segmentation
to classify Alzheimer’s using deep learning
Esraa H. Ali1, 2
, Sawsan Sadek1
, Zaid F. Makki3
1
Doctoral School of Sciences and Technology, Lebanese University, Beirut, Lebanon
2
Department of Computer Science, College of Science, Al-Nahrain University, Baghdad, Iraq
3
Al-Nahrain Center for Strategic Studies, Baghdad, Iraq
Article Info ABSTRACT
Article history:
Received Jan 19, 2024
Revised Mar 24, 2024
Accepted Jun 1, 2024
Brain damage and deficits in interactions among brain cells are the primary
causes of dementia and Alzheimer’s disease (AD). Despite ongoing research,
no effective medications have yet been developed for these conditions.
Therefore, early detection is crucial for managing the progression of these
disorders. In this study, we introduce a novel tool for detecting AD using non-
invasive medical tests, such as magnetic resonance imaging (MRI). Our
method employs fuzzy C-means clustering to identify features that enhance
image accuracy. The standard fuzzy C-means algorithm has been augmented
with fuzzy components to improve clustering performance. This enhanced
approach optimizes segmentation by extracting image information and
utilizing a sliding window to calculate center coordinates and establish a
stable group matrix. These critical features are subsequently integrated with a
two-phase watershed segmentation process. The resulting segmented images
are then used to train an optimal convolutional neural network (CNN) for AD
classification. Our methodology demonstrated a 98.20% accuracy rate in the
detection and classification of segmented MRI brain images, highlighting its
efficacy in identifying disease types.
Keywords:
Alzheimer’s disease
Convolutional neural network
Improved fuzzy C-means
Magnetic resonance imaging
Watershed segmentation
This is an open access article under the CC BY-SA license.
Corresponding Author:
Esraa H. Ali
Doctoral School of Sciences and Technology, Lebanese University
Al-Hadath District, Beirut, Lebanon
Email: esraa.ali@ul.edu.ib
1. INTRODUCTION
Alzheimer’s disease (AD) is the leading cause of dementia among older adults, characterized as a
mental health disorder that results in brain damage and impairs the ability to perform daily activities [1]. It is
a chronic neurodegenerative condition with an insidious onset and gradually worsening symptoms over time.
The etiology of AD remains unclear, and treatments are often expensive. In recent years, there has been a
significant focus on early d AD is the leading cause of dementia among older adults, characterized as a mental
health disorder that results in brain damage and impairs the ability to perform daily activities. It is a chronic
neurodegenerative condition with an insidious onset and gradually worsening symptoms over time. The
etiology of AD remains unclear, and treatments are often expensive. In recent years, there has been a significant
focus on early detection of this form of dementia by academics and researchers. The current global
demographic of individuals with dementia is estimated to be 47.5 million, projected to increase to 75 million
by 2030 [2], [3]. The advancement of digital neuroimaging techniques has enhanced the analysis of clinical
imaging data for diagnosing brain disorders. Techniques such as magnetic resonance imaging (MRI),
cerebrospinal fluid (CSF) analysis, single photon emission computed tomography (SPECT), and
fluorodeoxyglucose positron emission tomography (FDG-PET) are instrumental in identifying structural

Int J Artif Intell ISSN: 2252-8938 
Hybrid improved fuzzy C-means and watershed segmentation to classify Alzheimer’s … (Esraa H. Ali)
4081
changes in the brain. Since the early 1980s, the medical community has begun utilizing these advanced medical
imaging techniques to improve the quality of healthcare imagery [4], [5]. The evolution of traditional image
processing methods alongside machine learning (ML) and deep learning (DL) has led to significant
advancements in medical image analysis. Analytical image processing techniques are categorized into
registration, classification, detection, segmentation, and localization, with segmentation being a crucial step in
isolating the desired tissue or region of interest (RoI) from the collected images [6], [7].
The model architecture for diagnosing AD encompasses five key stages: data acquisition,
segmentation, registration, morphometry, and classification. Various standard datasets, such Alzheimer's disease
neuroimaging initiative (ADNI), international consortium for brain mapping (ICBM), minimal interval
resonance imaging in Alzheimer's disease (MIRIAD), Kaggle, open access series of imaging studies (OASIS),
Harvard Medical School, and others, are utilized to gather extensive data from morphological and anatomical
images. These images are essential for identifying abnormalities in the affected brain. Effective segmentation
and classification, particularly in MRI studies, necessitate a robust image pre-processing approach. In the
development of Alzheimer's detection systems, processes such as noise reduction, smoothing, skull stripping,
cropping, and normalization are indispensable. Through the registration process, images are aligned to a standard
reference area, facilitating intra-image and inter-image matching crucial for tracking disease progression and
identifying affected individuals [8]. Classification, the final stage, involves categorizing patients as normal or
exhibiting abnormalities. Artificial intelligence (AI), in conjunction with MRI, emerges as a promising method
for disease classification. The development and application of ML and DL are pivotal in crafting AI-based
classification algorithms that enhance outcomes, quality, and efficiency [9].
In various studies, convolutional neural network (CNN)-based learning has been found to lack
robustness, prompting the exploration of alternative methods to enhance performance. For MRI brain images,
a hybrid approach combining enhanced fuzzy C-means clustering with watershed segmentation (Ws) has been
utilized as a feature detection and extraction mechanism to delineate gray matter (GM), white matter (WM),
and CSF regions of the brain. Our literature review revealed that limited research has been conducted on
developing specialized CNN architectures for more effective AD.
The technique proposed in [10] initiates with a genetic algorithm (GA) for feature selection,
identifying the most informative subset of features. Fuzzy C-means (FCM) clustering is applied to this selected
subset. This approach reduces the dimensionality of the feature space, thereby rendering the classification
process via support vector machine (SVM) both more efficient and understandable. It notably enhances the
accuracy of early AD detection by accentuating the differentiation between AD and non-AD clusters. The
results underscore the efficacy of this technique in precisely identifying individuals at rick of Alzheimer's at
an early stage. To track AD progression, Sappagh et al. [11] introduced a multi-modal ensemble DL technique
that extracted both local and longitudinal information from each modality. Additionally, prior knowledge was
utilized to derive local features form MRI, positron emission tomography (PET), cognitive scores,
neuropathology, and ADNI assessments. Employing a combination of layered CNN-bidirectional long
short-term memory (BiLSTM), all gathered features were integrated for regression and classification tasks. A
multi-modal approach for automated hippocampus segmentation using 3D patches was detailed [12]. Utilizing
sMRI (T1) images from the ADNI dataset, a hybrid multi-task deep CNN and 3D DenseNet+softmax were
employed for AD classification. The model achieved an accuracy of 88%, sensitivity 86%, and an area under
the curve (AUC) of 92% [13]. This study presents a method for early detection of Alzheimer's using SVMs
trained on various texture descriptors, which aid in dimensionality reduction derived from MRI alongside
SVMs trained on markers obtained from ADNI. Different feature selection methods, each training a distinct
SVM, were applied to reduce the dimensionality of voxel-based features. Geetha and Pugazhenthi [14] suggests
a novel approach for Alzheimer's classification from MRIs using a fuzzy neural network (FNN). The wavelet
transformation (WT) is employed for image decomposition, with the discrete wavelet transform calculating the
output coefficient vectors. These generated features are then used to train FNN. A CNN model is proposed for
AD classification using MR images with hippocampus designated as the RoI [15]. The RoI is extracted through
an automated patch-based separation method that utilizes geometric values from the international consortium
for brain mapping (ICBM) standard. CNN was applied for dataset classification, demonstrating notable
performances. A novel methodology combining extreme learning and deep learning for AD classification is
introduced [16]. This approach evaluates two deep learning models for functional brain-network classification,
alongside an extreme learning machine (ELM) enhanced framework for learning deep regional-connectivity
and deep adjacent positional features. The construction of the brain network utilizes the Pearson correlation
coefficient. In summary, our review highlights three key findings:
− The majority of the literature were reviewed reported evaluation scores below 92%, whereas our study
achieved an exceptional performance of 98%. This significant advancement is attributed to our novel
segmentation and feature extraction model, which effectively reduces variable parameters while enhancing
computational speed.

 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 4, December 2024: 4080-4094
4082
− Our research introduces a potent classification architecture that utilize different kernel sizes to extract
essential features. By incorporating two smaller kernels (3×3 and 5×5), our model achieves robust training
and testing processes, thereby improving its performance and reliability.
− Furthermore, our study leverages a substantial dataset comprising 6400 brain images, categorized into four
distinct classes: mild, normal, moderate, and very mild. This extensive dataset offers a comprehensive
view across various stages of AD, thereby enriching the depth and diversity of our analysis.
2. METHOD
In this paper, we utilize the improved fuzzy C-means clustering (ImFCm) for segmenting brain tissue,
owing to its efficiency in segmenting homogeneous intensity regions of MRI images. We introduce a hybrid
approach combining ImFCm with Ws, achieving more effective results in accurately partitioning images and
enhancing classification performance. The outcomes of these two methods are then integrated into an optimized
CNN architecture, aiming to improve the accuracy and robustness of the AD detection system. The ADNI
dataset was employed to validate the findings, with approximately 6400 MRI brain images analyzed. These
images are annotated into four categories: mild, moderate, very mild, and normal. Figure 1 illustrates the block
diagram of the proposed approach. Each section of our method is explained in detail in the subsequent sections.
Figure 1. The proposed method
2.1. Pre-processing
Preprocessing is a technique of image enhancement that focus on both the data structure and
processing constraints. It encompasses the removal enhancement of the image to improve system performance.
Cropping is used to eliminate unnecessary components from an image. Additionally, converting the images to
grayscale, an essential step is performed. The contrast adaptive histogram equalization (AHE) filter and Bayes
wavelet transform (WT) are utilized to reduce noise, enhance brightness and contrast, and normalize the image.
Figure 1 demonstrates the preprocessing steps. This process aims to remove noise from MRI images. The DB3
wavelet is used for decomposing the image, and the noise standard deviation is considered when establishing
wavelet detail coefficient threshold. The type of wavelet applied is determined pywt.wavelist function, with
bior6.8 selected as the wavelet choice. Soft thresholding is implemented to find the optimal match for the
original image with additive noise. Contrast limited adaptive histogram equalization (CLAHE) is an advanced
version of AHE designed to prevent contrast over amplification. CLAHE operates on small sections of the

4083
image rather than the entire image, using a ClipLimit parameter to set contrast threshold. The initial value is
set at 3, with the tile grid size determining the number of tiles per row and column set to 8×8. This approach
applies a contrast filter by dividing the image into sections. The preprocessing stage concludes with cropping
and normalizing MRI images. Cropping is a technique in computer imaging used to remove irrelevant areas
and surroundings from images. Normalization is the process of reducing the intensity variation among pixel
values, marking the final phase of the preprocessing stage.
2.2. Segmentation process
Image segmentation methods encompass threshold-based, edge-based, region-based, matching-based,
clustering-based, fuzzy inference-based, and generalized principal component analysis techniques. Each
method offers advantages and limitations. Clustering is a method for dividing a collection of objects into
different groups, each known as a cluster. Members within each cluster exhibit high similarity in terms of
features, but the degree of similarity compared to members of other clusters is minimal. While many clustering
algorithms share foundational concepts, differences arise in how similarity or distance is measured and how
labels are assigned to categories within each cluster. Key strategies include fuzzy clustering, density-based
clustering, discriminative clustering, model-based clustering, and hierarchical clustering [17]. In our study, we
have combined ImFCm clustering with Ws to enhance both the accuracy and efficiency of image analysis.
2.2.1. Improved fuzzy C-means clustering (the proposed method)
In fuzzy clustering, unlike traditional clustering where each sample is assigned exclusively to one
cluster, a single sample can be associated with multiple clusters. The core principle behind fuzzy clustering is
that each element can be assigned to different clusters with varying degrees of membership [18]. FCM is a
widely recognized fuzzy clustering approach. Our objective is to optimize the following methodology [19]
using the FCM algorithm:
𝐽𝑚 = ∑ ∑ 𝑢𝑖𝑘
𝑚
𝑑𝑖𝑘
2
𝑛
𝑘=1 = ∑ ∑ 𝑢𝑖𝑘
𝑚
‖𝑥𝑘 − 𝑣𝑖‖2
𝑛
𝑘=1
𝑐
𝑖=1
𝑐
𝑖=1 (1)
where m is a positive integer that is greater than one. Moreover, uik is the kth data's level of membership in the
ith cluster, dik is the ratio of familiarity in the preceding n space, xk indicates the kth data, and vi is the ith
cluster's center. In our study, we aim to develop an enhanced and robust fuzzy C-means (FCM) clustering
technique, with modifications implemented in the following areas, as depicted in Figure 2.
Figure 2. Improved fuzzy C-means clustering

 ISSN: 2252-8938
4084
First part (initial parameters setting): the parameter settings were carefully chosen as follows; we
initially selected a cluster number of 4, as this count is considered optimal for segmenting MRI brain images
into four distinct regions, effectively highlighting key tissues within the human brain. The fuzziness degree
parameter was set to 2, providing increased flexibility in associating data with specific clusters, thereby
achieving a balance between sensitivity and robustness. To ensure adequate convergence, a limit of 100
iterations was chosen, this decision being informed by observations from convergence experiments. A neighbor
effect of 4 was selected to reflect the size of the sliding window used in image filtering, which facilitates the
computation of local features and captures spatial relationships. An epsilon threshold of 0.05 was established
for the convergence criteria, indicating a stringent convergence threshold due to its lower value. The kernel
size was determined to be 3, to aid in capturing spatial information during the image filtering process.
Second part: the distance window has been utilized to filter the image. A sliding window technique is
employed to traverse the entire image, aiming to identify stable groups. Initially, padding is created, equivalent
to half of the kernel size, to ensure the inclusion of image borders during the sliding process. The image mean
is calculated using this padding, and the cv2.copyMakeBorder function is employed to incorporate edges
during sliding. Subsequently, a sliding window algorithm is defined, specifying the neighbor effect and window
size as parameters, focusing on locating stable groups of pixels within clusters. The function for locating stable
groups operates by identifying stable pixels through a Gaussian filter, where the filter's values are less than or
equal to the square root of the nan mean for the power difference of Gaussian values and window size. The
result of this phase is a filtered image, which will be further utilized in the third part to compute fuzziness.
Third part: the histogram of the image is determined using the CLAHE filter. It is to enhance image
contrast and calculate the intensity distribution of the enhanced image. This phase is instrumental in
determining the centroid of the cluster.
Fourth part: this segment encompasses several critical functions. It begins with the initialization of the
membership function, which determines the degree to which the pixels of the image belong to each cluster.
This is followed by the function for computing the centroids of the clusters, which involves the division of the
numerator by the denominator. The numerator is the sum of the product of the degree of fuzziness, the
histogram, and the intensity, each raised to the power of membership. Conversely, the denominator is the
summation of the histogram values raised to the power of membership. The final step in this part is the
computation of weights, a process reliant on the centroid computation function. This involves dividing the
numerator by the denominator, where the numerator calculates the absolute differences between the intensity
raised to a power and the computed cluster centroids. The denominator, on the other hand, sums the absolute
differences, each raised to the power of the fuzziness degree. Algorithm 1 shows the detailed process. The
block diagram in Figure 2 shows the process of utilizing the improved fuzzy C-means clustering.
Algorithm 1: ImFCM
Step 1: Initialize the following parameters:
- Number of bits. Number of clusters.
- Degree of fuzziness.
- Maximum iteration count.
- Epsilon threshold for convergence check.
Step 2: Image Filtering Procedure:
- Generate a padded image using a sliding window with dimensions (kernel_size/2,
kernel_size/2).
- Compute the mean based on the padded mask.
- Pad the resulting mean image to create borders.
- Utilize a sliding window to account for neighbor effects and kernel size.
- Determine center coordinates using the spatial distance window with Minkowski
distance:
Des_win = ((abs (win_size_y - center_coordinate_y)) ** p + abs ((win_size_y -
center_coordinate_y) ** p)) ** (1/p), where p = 2.
- Identify the stable group matrix using a Gaussian filter.
- Obtain the final filtered image using the formula:
Final_image = sum (weighted_coefficients * old_window) / sum (weighted_coefficients)
- Perform CLAHE.
Step 3: Weight Initialization: Initialize a two-dimensional matrix based on the number of
clusters and gray levels to compute weights.
Step 4: Compute Cluster Centroids:
- Calculate the X and Y values as follows:
- X = sum (histogram * number of gray levels) * power (weight * number of fuzziness)
- Y = sum (histogram) * power (weight * number of fuzziness)
- Z = X / Y
Step 5: Weight Computation Method:
- Set power = -2 / number of fuzziness.
- Calculate the X value using the formula: X = (gray levels - centroid values) * power.

4085
- Compute Y as: Y = sum (gray levels - centroid values) * power. Determine Z as:
Z = X / Y.
Step 6: Check Convergence: Determine whether the absolute maximum value of (step 5 - step 2)
is less than the epsilon threshold. If so, stop; otherwise, proceed to step 4.
2.2.2. Watershed segmentation
In Ws, an image is conceptualized in three dimensions, with the (x, y) coordinates correspond to the
spatial axes and the intensity represented along the z-axis. This approach treats an image as if it were a
topographical landscape, with the intensity of each pixel analogous to elevation levels. Consequently, each
intensity level is associated with a distinct elevation plane on this landscape. Utilizing this topographical
metaphor, points within the image are categorized into three segemts: regional minima, catchment basins, and
watershed lines. Catchment basin are areas were, hypothetical, a droplet of water would coverage towards a
single regional minimum. Watershed lines, conversely, mark the boundaries where a droplet of water could
potentially be drawn towards multiple regional minima, effectively delineating the division between distinct
catchment basins [11].
Consider the M1, M2, … MR regional minima of an image g(x, y). Let T[n] represent an array of
points beneath the horizontal axis with a value of n, where n ranges from the image's least to greatest intensity.
This may be stated mathematically as follows:
T[𝑛] = {(𝑠, 𝑡)|𝑔(𝑠, 𝑡) < 𝑛} (2)
Cn(Mi) indicate a collection of regions in the catchment basin that are poured at plane n that are related with
the region minimum Mi. This could possibly be used to compute it by:
𝐶𝑛(𝑀𝑖) = 𝐶(𝑀𝑖)𝑇[𝑛] (3)
C(Mi) is the set of catchment basin points linked with the regional minimum Mi. The union of all flooded
catchment basins at a certain stage n represented in C[n]:
𝐶[𝑛] = [𝐶𝑛(𝑀𝑖)] (4)
Algorithm 2 introduces the steps of this technique. Figure 3 illustrate the watershed method in a block diagram.
Algorithm 2: Watershed segmentation
- Utilize OTSU’s binarization filter to estimate the objects present in the image.
- Apply morphological opening to eliminate any white noise present in the image, and perform
morphological closing to address small holes within the objects.
- Employ the dilate method to create a separation between the background and the image.
- Utilize distance transform and thresholding techniques to isolate the foreground from the background.
- Determine the unknown areas by subtracting the foreground from the background. These areas lacking
clarity will be assigned zero values in the markers.
- Label the regions of the foreground using the connected components method as markers, and increment
them by one to ensure all background regions are marked as ones.
- Employ the distance values obtained from step 5 and the markers from step 6 as input parameters for the
watershed method to generate the final segmentation map.
Figure 3. Watershed segmentation

 ISSN: 2252-8938
4086
2.3. Post-processing
Gamma correction is a technique used for data augmentation that entails adjusting the gamma value
to modify the image intensity. Gamma represents a non-linear function that can be applied to either encode or
decode the brightness or intensity of an image [20]. Gaussian noise, another form of distortion, is introduced
into the image through random values drawn from a Gaussian distribution. Since noise encountered during
image acquisition and preprocessing can escalate, applying Gaussian noise to a raw image may help the model
become more resilient to variation in image quality [21]. Our study advocates for the use of these two methods
as a post-processing measure for datasets that may suffer from loss in contrast and brightness, as well as to
introduce a slight blurring effect to smooth transitions in pixel value and soften the image edges. The marker-
controlled Ws technique is applied following the ImFCm twice with a marker value range of [10-90] to capture
images highlighting internal brain features. This process is then repeated with a marker value range of
[10-200] to obtain images showcasing external brain details. The culmination of this process involves
combining all three images to produce the final enhanced image. Post-processing steps, including gamma
correction and Gaussian blur, are subsequently performed to further refine the images, as depicted in Figure 4.
Figure 4. Segmentation stage
2.4. Classification
CNN, a cornerstone of the neural network framework, encompasses numerous layers within its
architecture and has gained significant prominence in various image processing applications, notably in object
recognition [22] and image classification, where it has yielded promising outcomes. Previous research indicates
the feasibility of directly inputting images into a CNN network to extract features for image categorization.
The architecture of a CNN comprises several fundamental components, including convolutional layers,
SoftMax layers, pooling layers, non-linear activation functions such as the rectified linear unit (ReLU), and
fully connected layers (FC) [23], [24]. CNNs operate based on the intensities of images, utilizing dimensions
such as width, height, and depth to represent the input image intensities. The processing begins from the top
left corner of the image and progresses to the right. As the filter moves from the top to the bottom of the input
volume, it changes, with each left-to-right movement constituting a stride. The complexity of the stride is
augmented by the number of steps it encompasses. ReLU serves as an efficient activation function by
converting negative pixel values to zero [25]. Following the convolution process, the size of the hidden layer
becomes significantly large, necessitating the use of a pooling or sub-sampling layer to reduce computational
complexity. Pooling can be categorized into two types: maximum and average [26]. Within the context of
pooling, let y = yij represent the matrix.
𝑅𝑒𝐿𝑈(𝑌) = max(0, 𝑌) (5)
As noticed in (6), max pooling is the process that selects the most significant component in y as the output
𝑥 = max(𝑦) (6)
Post-Processed Image
ImFCm Ws with marker [10-90] Ws with marker [10-200]
Pre-processed Image
Blended Image

4087
Our study analyzed MRI brain images from AD patients, organizing the dataset into four distinct
categories: normal, mild, moderate, and very mild. Figure 5 illustrates a proposed CNN architectural model
comprising 13 layers. This model includes four convolutional layers, with each pair of convolutions followed
by a pooling layer and dropout layers to ensure regularization. The architecture concludes with a fully
connected layer and a classifier layer. The original images, measuring 176×208 pixels, are resized to 200×200
pixels before being input into the CNN model. The filter size is varied across the CNN layers to effectively
identify features. The batch size is set to 32, and the model undergoes training over 50 epochs. Upon completion
of all epoch cycles, the CNN selects the model iteration with the highest performance metrics for classification
purposes. The final classification is then applied to the test set to determine the accuracy rate.
Figure 5. CNN model layers
Table 1 outlines the internal architecture of the CNN model, detailing the specific layers and
configurations used within the model. Table 2 lists the hyperparameters applied during the model’s training
and optimization processes. For optimization, The Adam algorithm is utilized, with a learning rate of 0.001 set
for the entire training phase.
Table 1. CNN model
Model layers Image volume Filters Size of filter Pooling win size var.
conv2d (Conv2D) (200, 200) 32 5x5 2x2 2432
conv2d_1 (Conv2D) (200, 200) 32 5x5 2x2 25632
MaxPooling2D (100, 100) 32 2x2 0
Dropout (100, 100) 32 0
conv2d_2 (Conv2D) (100, 100) 64 3x3 18496
conv2d_3 (Conv2D) (100, 100) 64 3x3 36928
max_pooling2d_1 (50, 50) 64 2x2 0
dropout_1 (50, 50) 64 0
flatten (Flatten) (None, 160000) 0
dense (Dense) (None, 256) 40960256
dropout_2 (None, 256) 0
dense_1 (Dense) (None, 4) 1028
Table 2. Values of hyper-parameters
Hyper-parameters Value
Split data 3840 train, 1281 validate
Dropout 0.3, 0.3, 0.5
Batch size 32
Learning rate 0.001
Num. of epochs 50

 ISSN: 2252-8938
4088
3. RESULTS AND DISCUSSION
3.1. Experimental dataset
The MRI images utilized for this research were sourced from the ADNI database. A total of 6400
ADNI samples were selected for analysis after excluding certain samples with incorrect information. The
dataset comprises 3140 normal samples, 896 early mild cognitive impairment samples, 64 moderate cognitive
impairment samples, and 2240 severe cognitive impairment samples. These images are in JPEG format with a
resolution of 176×208 pixels.
3.2. Evaluation metrics
3.2.1. Improved fuzzy C-means clustering
The evaluation criteria used to assess computational complexity focus on how efficiently our method
performs in terms of computational resources used and the quality of results obtained. It involves evaluating
the time and scalability of our method when applied to extensive datasets. Consideration is given to techniques
or optimizations that could enhance the algorithm's efficiency without compromising accuracy. The used
equation to compute the efficiency of our proposed method is depicted (7):
𝑑 = ∑ ∑ |𝑈𝑖𝑗 − 𝑜𝑙𝑑_𝑢𝑖𝑗|
𝑛
𝑗=1
𝑐
𝑖=1 (7)
where 𝑑 denotes the total absolute difference. 𝑐 signifies the number of clusters. 𝑛 represents the number of
data points. 𝑈𝑖𝑗 reflects the membership value of data point j in cluster i during the current iteration. 𝑜𝑙𝑑_𝑢𝑖𝑗
indicates the membership value of data point j in cluster i during the previous iteration. The equation calculates
the sum of the absolute differences between corresponding elements of the current and prior membership
matrices across all clusters and data points. This measurement can act as an indicator of convergence or change
between successive iterations within an optimization algorithm, such as the fuzzy C-means clustering
algorithm.
3.2.2. Convolutional neural network
The analysis of ML and DL recognition platforms, aimed at evaluating their capability to accurately
diagnose AD, relies on several performance metrics, including accuracy (Acy), sensitivity (Sny)/recall,
precision (Prn), and the F1 score. Each of these performance indicators offers different insights into the
proposed model's effectiveness. The primary measure for evaluating the classification system is accuracy,
which is calculated by dividing the number of correct predictions by the total number of predictions made.
Mathematically, it can be expressed as (8):
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑝+ 𝑇𝑁
𝑇𝑝+𝐹𝑝+𝑇𝑁+𝐹𝑁
(8)
where TP and TN are true positive and true negative respectively. FP, FN are false positive and false negative,
respectively. Sensitivity and specificity are logically specified as (9):
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =
𝑇𝑝
𝑇𝑝+𝐹𝑁
(9)
The sensitivity metric acts as an indicator of the effectiveness in detecting AD patients, reflecting the
model's ability to correctly identify those who are truly affected by the disease. Precision measures the
reliability of the diagnosis, representing the proportion of individuals identified by the system as having the
disease who are indeed seriously impacted by it. This can be described as (10):
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑝
𝑇𝑝+𝐹𝑝
(10)
The F1 score of the simulation is described as the average of the sensitivity and accuracy.
𝐹1 𝑠𝑐𝑜𝑟𝑒 = 2 × (
𝑆𝑛𝑦 × 𝑃𝑟𝑛
𝑆𝑛𝑦×𝑃𝑟𝑛
) (11)
3.3. Experimental results
3.3.1. Comparison results traditional fuzzy C-means and improved fuzzy C-means clustering
Upon completion of the ImFCm algorithm, convergence was achieved by the 35th iteration, thereby
exceeding the preset maximum iteration threshold. The cost value experienced a significant reduction,
descending from 385.01 to 0.049 as the iterations advanced. This decline signifies the algorithm’s convergence

4089
towards an optimal solution, highlighting the efficiency of our method in achieving results in a shorter
timeframe. In comparison, the conventional FCM algorithm attained convergence at the 70th iteration, intiating
with a cost value of 4907.9, and thus requiring more time relative to the proposed method. Figure 6 depicts the
cost values for five different images alongside the duration taken by both the FCM and ImFCm algorithms.
Figure 6. Cost and time for FCM and ImFCm
3.3.2. Convolutional neural network results with traditional fuzzy C-means
Traditional FCM clustering were employed for segmentation, which was then input into CNN for AD
classification. Figure 7 illustrates the training and validation results from the segmented MRI brain dataset.
The classification results demonstrated a test accuracy of 91% achieved over 50epochs. One of the curves
shows a red line representing training loss and a blue line for validation loss, while another curve illustrates a
red line for validation accuracy and a blue line for accuracy. According to this technique, the figure indicates
that both accuracy and validation accuracy converged after 10 epochs. The validation accuracy reached 0.88
and remained constant from epoch 14 through to epoch 50. By epoch 50, the loss decreased to 0.04, and the
validation loss reached 0.6.
0
50
100
150
200
250
300
350
400
450
1 4 7 10 13 16 19 22 25 28 31 34 37
COST
Iteration
ImFCm
Time:
Image1: 4.322
Image2: 4.346
Image3: 4.645
Image4: 4.973
Image5: 4.426
0
10000
20000
30000
40000
50000
60000
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70
COST
Iteration
FCM
Time:
Image1: 23.723
Image2: 24.426
Image3: 25.380
Image4: 26.312
Image5: 27.255

 ISSN: 2252-8938
4090
Figure 7. The accuracy and loss curves of classification for traditional FCM
Table 3 displays the classification report of the training model for each class using this segmentation
technique. For the classes mild, normal, moderate, and very mild, the precisions were 91%, 100%, 91%, and
90%, respectively. The recall rates for these classes were 79%, 89%, 95%, and 90% in that order. The F1-
scores for the classes are 85%, 94%, 93%, and 90%, respectively.
Table 3. Classification Report of Traditional FCM
Class Name Precision Recall F1-Score
Mild 0.91 0.79 0.85
Normal 1.00 0.89 0.94
Moderate 0.91 0.95 0.93
Very mild 0.90 0.90 0.90
Accuracy 91% for predictions
3.3.3. Convolutional neural network results with improved fuzzy C-means clustering
The outcomes of the proposed classification process, as applied to both the training and validation
sets, are illustrated through accuracy and loss curves in Figure 8. The figure reveals that convergence of the
training approach's accuracy and loss was observed after 10 iterations, indicating high training and testing
accuracy. The training accuracy achieved is in the vicinity of 99%, accompanied by a loss of approximately
3%, whereas the validation accuracy approaches 98%, with a loss near 6%. Analysis of the figures demonstrates
that the discrepancy between training accuracy and validation accuracy, as well as between training loss and
validation loss, is minimal. Consequently, 50 epochs have been deemed suitable for the training and verification
of our model. Figure 8 further indicates that after the initial 10 epochs, the accuracy for both training and
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
LOSS
EPOCHS
Training Loss Validation Loss
0
0.2
0.4
0.6
0.8
1
1.2
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
ACCURACY
EPOCHS
Training Accuracy Validation Accuracy

4091
validation has stabilized. Based on these results, it can be concluded that the method we proposed has enhanced
the model's efficiency, enabling it to be trained and validated within fewer epochs.
Figure 8. The accuracy and loss curves of the proposed classification results
By integrating two powerful segmentation techniques, the updated model reveals more detailed
features. The attributes derived from these hybrid methods are varied; by combining them, the features become
more robust, enhancing the classification phase. Table 4 displays the classification report of the training model
for each class. For the classes mild, normal, moderate, and very mild, the precisions were 97%, 100%, 99%,
and 97%, respectively. The recall rates for these classes were 99%, 100%, 98%, and 99%, in that order. The
F1-scores for the classes are 98%, 100%, 98%, and 98%, respectively. The outcomes of the adapted model
used in this study generally demonstrate exceptional performance, indicating that employing advanced MRI
segmentation techniques to enhance AD diagnostic classification performance is beneficial. After
comprehensive training, the system undergoes evaluation using a testing set, which consists of images that
were not exposed to the system during the training phase. Employing our recommended segmentation
technique, the CNN model achieves an accuracy of 98.98% and demonstrates efficient performance on MRI
images.
Table 4. Classification report of the proposed method
Class name Precision Recall F1-Score
Mild 0.97 0.99 0.98
Normal 1.00 1.00 1.00
Moderate 0.99 0.98 0.98
Very mild 0.97 0.99 0.98
Accuracy 98.20% for predictions
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
LOSS
EPOCHS
Training Loss Validation Loss
0
0.2
0.4
0.6
0.8
1
1.2
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
ACCURACY
EPOCHS
Training Accuracy Validation Accuracy

 ISSN: 2252-8938
4092
3.4. Comparative examination from diverse researches models
Our proposed methodology distinguishes itself through a comparative analysis with existing research
efforts aimed at early AD detection, utilizing a detailed four-class classification framework, as detailed in
Table 5. The suggested ImFCm-Ws-CNN model exhibits notable advancements in accuracy, precision, and
F1-score metrics. It achieves an impressive 96.25% accuracy, 98.0% precision, and a 97% F1 score, surpassing
the performance of previous studies. This comparative assessment underscores the advantages of incorporating
ImFCm-Ws techniques into CNN models, highlighting enhanced diagnostic capabilities without sacrificing
performance metrics.
Table 5. A summary of new studies implementing techniques using DL
Ref. Image Dataset
Extracted
Features
Classifier Acc precision Recall F1 Others
[27] T1-MRI MIRIAD --- CNN 89.0 89.0 89.0 89.0 AD Vs Non
[28] MRI OASIS --- HT-MGWRO 93.16 90.74 94.23 92.45 AD Vs Non
[29] MRI+EEG ADNI+
Rowan
Univ
Texture
properties
of an image
Hybrid
CNN+DBN
92.50 --- 90.89 --- AD Vs Non
[30] MRI ADNI FreeSurfer DNN 85.19 76.93 72.73 74.77 Multiclass
[12] SMRI (T1) ADNI 3D Patches Hybrid multi-
task deep
CNN and
DenseNet+sof
tmax
88.90 --- 88.60 --- AD Vs Non
[31] MRI ADNI Bag of
Features
(BoF)
SVM 93 ---- ---- ---- Multiclass
[32] MRI OASIS Statistical
measures of
the mean
and
standard
deviation
Hybrid
AlexNet+
SVM
94.0 93.0 97.0 --- Multiclass
[11] MRI, PET,
Cognitive
scores,
Neuropathology,
assessment
ADNI Local and
longitudinal
features
Stacked
CNN-
BiLSTM
92.62 --- 98.42 --- AD
progression
Proposed MRI ADNI ImFCm-Ws CNN 98.20 98 98 98 Multiclass
4. CONCLUSION
This study introduces a novel methodology for the precise identification of AD progression through
the integration of feature extraction using the ImFCM-Ws technique with an optimized CNN architecture.
Utilizing the standardized ADNI dataset, our model effectively categorizes the different stages of AD. The
evaluation revealed that the CNN model, enhanced with ImFCM-Ws features, outperformed existing methods,
achieving an exceptional accuracy of 98.20%. Through meticulous feature extraction, our approach accurately
identifies brain regions associated with Alzheimer's pathology, providing invaluable assistance to healthcare
professionals in evaluating the disease’s severity based on levels of dementia. By allowing the data to inform
our analysis, our findings underscore the significance of our methodology in improving the detection and
diagnosis of AD. On the other hand, several limitations must be introduced despite the results and promising
contributions of our study. One of these limitations, our approach relies on the ADNI dataset, which, although
comprehensive, may not include all data sources in AD. This limitation could introduce bias if inaccuracies
have occurred in generalizing our model to specific populations or cases with specific characteristics. To
prevent this limitation, we suggest for future to add multimodal data sources such as neuroimaging scans,
genetic information, and clinical assessments that could refine our model to capture additional information to
detect this disease.
5. FUTURE WORK
Our study employs a synergistic approach, combining ImFCm with Ws for effective feature
extraction, and leveraging CNN for classification. While our findings illuminate the effectiveness of this
methodology in detecting AD, several areas warrant further investigation. Future research could focus on
optimizing the parameters and algorithms of ImFCm and Ws to enhance the accuracy and efficiency of feature

4093
extraction. Additionally, the exploration of innovative fusion techniques for integrating multimodal
neuroimaging data may significantly enhance the diagnostic accuracy and robustness of the classification
model. Moreover, the application of advanced ML and DL architectures beyond CNN, such as recurrent neural
networks (RNNs) or graph neural networks (GNNs), could provide novel perspectives on AD diagnosis and
progression monitoring. These prospective research avenues aim to propel forward the domain of
neuroimaging-based AD detection, contributing towards the creation of more precise and dependable
diagnostic tools.
REFERENCES
[1] Alzheimer’s Dementia, “2020 Alzheimer’s disease facts and figures,” The Journal of the Alzheimer's Association, Mar. 2020, pp.
391–460, doi: 10.1002/alz.12068.
[2] G. Liu, Q. Lin, N. N. Xiong, and X. Wang, “Unsupervised denoising feature learning for classification of corrupted images,” Big
Data Research, vol. 27, Feb. 2022, doi: 10.1016/j.bdr.2021.100305.
[3] M. Liu, M. Zhou, T. Zhang, and N. Xiong, “Semi-supervised learning quantization algorithm with deep features for motor imagery
EEG Recognition in smart healthcare application,” Applied Soft Computing, vol. 89, Apr. 2020, doi: 10.1016/j.asoc.2020.106071.
[4] D. Li et al., “CARM: Confidence-aware recommender model via review representation learning and historical rating behavior in
the online platforms,” Neurocomputing, vol. 455, pp. 283–296, Sep. 2021, doi: 10.1016/j.neucom.2021.03.122.
[5] C. Wu, C. Luo, N. Xiong, W. Zhang, and T.-H. Kim, “A greedy deep learning method for medical disease analysis,” IEEE Access,
vol. 6, pp. 20021–20030, 2018, doi: 10.1109/ACCESS.2018.2823979.
[6] G. Litjens et al., “A survey on deep learning in medical image analysis,” Medical Image Analysis, vol. 42, pp. 60–88, Dec. 2017,
doi: 10.1016/j.media.2017.07.005.
[7] L. Sun, W. Shao, M. Wang, D. Zhang, and M. Liu, “High order feature learning for multi-atlas based label fusion: application to
brain segmentation with MRI,” IEEE Transactions on Image Processing, vol. 29, pp. 2702–2713, 2020, doi:
10.1109/TIP.2019.2952079.
[8] E. H. Ali, S. Sadek, and Z. F. Makki, “A review of AI techniques using MRI brain images for Alzheimer’s disease detection,” in
2023 Fifth International Conference on Advances in Computational Tools for Engineering Applications (ACTEA), Jul. 2023, pp.
76–82, doi: 10.1109/ACTEA58025.2023.10194002.
[9] G. Currie, K. E. Hawk, E. Rohren, A. Vial, and R. Klein, “Machine learning and deep learning in medical imaging: intelligent
imaging,” Journal of Medical Imaging and Radiation Sciences, vol. 50, no. 4, pp. 477–487, Dec. 2019, doi:
10.1016/j.jmir.2019.09.005.
[10] I. Cherian, M. Alate, A. B. Desai, M. R. Prajna, and D. Rawat, “Early detection of alzheimer’s disease using fuzzy C-means
clustering and genetic algorithm-based feature selection from PET scan,” International Journal of Intelligent Systems and
Applications in Engineering, vol. 12, no. 3s, pp. 452–463, 2023.
[11] S. E.-Sappagh, T. Abuhmed, S. M. Riazul Islam, and K. S. Kwak, “Multimodal multitask deep learning model for Alzheimer’s
disease progression detection based on time series data,” Neurocomputing, vol. 412, pp. 197–215, Oct. 2020, doi:
10.1016/j.neucom.2020.05.087.
[12] M. Liu et al., “A multi-model deep convolutional neural network for automatic hippocampus segmentation and classification in
Alzheimer’s disease,” NeuroImage, vol. 208, Mar. 2020, doi: 10.1016/j.neuroimage.2019.116459.
[13] L. Nanni, S. Brahnam, C. Salvatore, and I. Castiglioni, “Texture descriptors and voxels for the early diagnosis of Alzheimer’s
disease,” Artificial Intelligence in Medicine, vol. 97, pp. 19–26, Jun. 2019, doi: 10.1016/j.artmed.2019.05.003.
[14] C. Geetha and D. Pugazhenthi, “Classification of alzheimer’s disease subjects from MRI using fuzzy neural network with feature
extraction using discrete wavelet transform,” Biomedical Research, pp. S14–S21, 2018, doi: 10.4066/biomedicalresearch.29-16-
2319.
[15] B.-K. Choi et al., “Convolutional neural network-based MR image analysis for Alzheimer’s disease classification,” Current Medical
Imaging Formerly Current Medical Imaging Reviews, vol. 16, no. 1, pp. 27–35, Jan. 2020, doi:
10.2174/1573405615666191021123854.
[16] X. Bi, X. Zhao, H. Huang, D. Chen, and Y. Ma, “Functional brain network classification for Alzheimer’s disease detection with
deep features and extreme learning machine,” Cognitive Computation, vol. 12, no. 3, pp. 513–527, May 2020, doi: 10.1007/s12559-
019-09688-2.
[17] T. Singh, N. Saxena, M. Khurana, D. Singh, M. Abdalla, and H. Alshazly, “Data clustering using moth-flame optimization
algorithm,” Sensors, vol. 21, no. 12, Jun. 2021, doi: 10.3390/s21124086.
[18] E. H. Ruspini, J. C. Bezdek, and J. M. Keller, “Fuzzy clustering: A historical perspective,” IEEE Computational Intelligence
Magazine, vol. 14, no. 1, pp. 45–55, Feb. 2019, doi: 10.1109/MCI.2018.2881643.
[19] R. Xu and D. WunschII, “Survey of clustering algorithms,” IEEE Transactions on Neural Networks, vol. 16, no. 3, pp. 645–678,
May 2005, doi: 10.1109/TNN.2005.845141.
[20] A. Zhong et al., “Deep metric learning-based image retrieval system for chest radiograph and its clinical applications in COVID-
19,” Medical Image Analysis, vol. 70, 2021, doi: 10.1016/j.media.2021.101993.
[21] I. Sirazitdinov, M. Kholiavchenko, R. Kuleev, and B. Ibragimov, “Data augmentation for chest pathologies classification,” in 2019
IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), Apr. 2019, pp. 1216–1219, doi:
10.1109/ISBI.2019.8759573.
[22] L. J. Belaid and W. Mourou, “Image segmentation: A watershed transformation algorithm,” Image Analysis & Stereology, vol. 28,
no. 2, pp. 93–102, May 2011, doi: 10.5566/ias.v28.p93-102.
[23] N. A. Mohammed, M. H. Abed, and A. T. A. -Salih, “Convolutional neural network for color images classification,” Bulletin of
Electrical Engineering and Informatics, vol. 11, no. 3, pp. 1343–1349, Jun. 2022, doi: 10.11591/eei.v11i3.3730.
[24] K. O’Shea and R. Nash, “An introduction to convolutional neural networks,” arXiv-Computer Science, pp. 1-11, Nov. 2015.
[25] R. Poojary, R. Raina, and A. Kumar Mondal, “Effect of data-augmentation on fine-tuned CNN model performance,” IAES
International Journal of Artificial Intelligence (IJ-AI), vol. 10, no. 1, pp. 84–92, Mar. 2021, doi: 10.11591/ijai.v10.i1.pp84-92.
[26] F. Fachruddin, S. Saparudin, E. Rasywir, and Y. Pratama, “Network and layer experiment using convolutional neural network for
content based image retrieval work,” TELKOMNIKA (Telecommunication Computing Electronics and Control), vol. 20, no. 1, pp.
118–128, Feb. 2022, doi: 10.12928/telkomnika.v20i1.19759.
[27] K. D. Silva and H. Kunz, “Prediction of Alzheimer’s disease from magnetic resonance imaging using a convolutional neural

 ISSN: 2252-8938
4094
network,” Intelligence-Based Medicine, vol. 7, 2023, doi: 10.1016/j.ibmed.2023.100091.
[28] N. Bharanidharan and R. Harikumar, “Modified grey wolf randomized optimization in dementia classification using MRI images,”
IETE Journal of Research, vol. 68, no. 4, pp. 2531–2540, Jul. 2022, doi: 10.1080/03772063.2020.1715852.
[29] A. Shikalgar and S. Sonavane, “Hybrid deep learning approach for classifying Alzheimer disease based on multimodal data,” in
Computing in Engineering and Technology, 2020, pp. 511–520, doi: 10.1007/978-981-32-9515-5_49.
[30] R. Prajapati, U. Khatri, and G. R. Kwon, “An efficient deep neural network binary classifier for Alzheimer’s disease classification,”
in 2021 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), Apr. 2021, pp. 231–234,
doi: 10.1109/ICAIIC51459.2021.9415212.
[31] H. Guo and Y. Zhang, “Resting state fMRI and improved deep learning algorithm for earlier detection of Alzheimer’s disease,”
IEEE Access, vol. 8, pp. 115383–115392, 2020, doi: 10.1109/ACCESS.2020.3003424.
[32] B. A. Mohammed et al., “Multimethod analysis of medical records and MRI images for early diagnosis of dementia and Alzheimer’s
disease based on deep learning and hybrid methods,” Electronics, vol. 10, no. 22, Nov. 2021, doi: 10.3390/electronics10222860.
BIOGRAPHIES OF AUTHORS
Esraa H. Ali holds both an M.Sc. and B.Sc. degree from Al-Nahrain University,
Iraq, which were attained in 2009. Presently, she serves as an assistant lecturer at the
Department of Computer Science within the College of Science at Al-Nahrain University,
located in Baghdad, Iraq. Her expertise lies in the field of image processing. Her research
interests encompass image processing, machine learning, and audio cryptography. To date,
she has authored approximately 7 papers featured in international journals and conferences.
She can be contacted at email: esraa.ali@ul.edu.ib or esraa.hussien@nahrainuniv.edu.iq.
Sawsan Sadek holds a Ph.D. in electronics from the University of Science and
Technology in Lille, France in 1996 and a Bachelor's degree of Physics from the Lebanese
University in 1990. Currently she is a Professor at the Lebanese University, Faculty of
Technology in Saida, Lebanon, Department of Communications and Information Networks.
She headed the department for 5 years from 2009 to 2014. She organized the thirteenth edition
of the IEEE MMS2013, and launched the first global conference on sensor networks and
smart technologies in Beirut IEEE SENSET2017 and issued a special edition of the scientific
journal ALOG Springer. She has many research papers in international scientific journals and
conferences and supervised many doctoral theses in cooperation with French universities.
She can be contacted at email: sawsansadek@yahoo.fr or dsadek@ul.edu.lb.
Dr. Zaid F. Makki holds Ph.D. Degree in Computer Information Systems (CIS)
form the Lebanese University in 2019. He also received his master’s degree of Computer
Information Systems (CIS) Middle East University, Jordan, in 2012. He also holds a master’s
degree in Educational Administration and Leadership, Middle East University, Jordan, in
2015. He is currently a Governance Advisor in the General Secretariat of the council of
Ministers and he is the Head of the digital transformation team Office of the Prime Minister
Scientific and Technological Advisor, he also occupies on behalf of National Security
Advisory Director General of the National Spatial Data Center of the Iraqi government. He
can be contacted at email: zaidfj@coeng.uobaghdad.edu.iq.

Hybrid improved fuzzy C-means and watershed segmentation to classify Alzheimer’s using deep learning

More Related Content

Similar to Hybrid improved fuzzy C-means and watershed segmentation to classify Alzheimer’s using deep learning (20)

More from IAESIJAI (20)

Recently uploaded (20)

Hybrid improved fuzzy C-means and watershed segmentation to classify Alzheimer’s using deep learning