1. Introduction

To advance the efficiency and accuracy of the medical diagnostic system, especially those that are distributed in complex areas (e.g., brain, skin, lung, and blood cancer classification), several live line diagnostic models (Bengio et al. [1]; Tiwari et al. [2]; Bhatt et al. [3]) work with image processing. The effectiveness of the detection and accuracy of the multidisciplinary medical data diagnostic system depends largely on the quality of the included images captured by a few techniques (Cetin et al. [4]) such as Magnetic Resonance Imaging (MRI) for brain data analysis, dermoscopic for skin, microscopic for blood leukemia, computed tomography (CT) scan images for lungs, etc. However, due to the uncontrollable lighting conditions and lots of noise availability during capturing, the illumination distributed on the surface of the medical images remains uneven, especially when backlight or fixed lighting conditions affect the diagnostic model (Bonabeau et al. [5]; Banks et al. [6]; Siva Raja and Rani [7]). It also leads to a comparison of the global low and local image and weak data in the black region and each aspect of the image plays a significant function in the examination of medical data. Therefore, preliminary processing is an important part of medical image classification because it plays an important role in computer-assisted medical diagnostic programs in different systems. Due to the position variability of the targeted regions, traditional hybrid segmentation technique such as Fuzzy Competitive Learning based Counter Propagation Network (FCPN) still works better than soft computing techniques [8]. Moreover, image classification is considered the most important process for medical imaging as it extracts the region of interest (ROI) from various data by semiautomatic or automatic process (Kennedy and Eberhart [9]). It classifies the image in areas based on a specific definition, such as the segmentation of damaged body parts or tissues in the local medical diagnostic system and the acquisition of boundaries and classification. There are some samples of medical image segmentations shown in Figure 1.

For medical experts and researchers alike, performing good medical image segmentation is a difficult undertaking (Karaboga [10]). Many researchers, on the other hand, have previously attempted to create an effective algorithm for medical image segmentation in order to aid in the identification of various disorders and diseases. Therefore, in this research work, we present a comparative medical image segmentation framework using swarm intelligence with Convolutional Neural Networks (CNN). Mainly, three different medical image segmentation mechanisms are considered and studied, namely, Traditional Segmentation (Fuzzy C-means (FCM) and K-means), Swarm Intelligence (Particle Swam Optimization (PSO)), and CNN-based segmentation with traditional approaches, as shown in Figure 2 with subclasses of architectures.

Used traditional segmentation techniques are the type of unsupervised machine learning and useful to find out groups or different patterns in medical data. In general terms, it’s an unsupervised activity that divides unstructured data into several groups based on their similarity and dissimilarity (Yang, 2009 [11]). The motivation behind the proposed scenario is given in further sections.

2. Literature Survey

Segmentation is widely used in various sectors such as split geographical regions, fruit from trees, flood for damage reports, recognition of traffic signs, and road collapses. Chouhan et al. [12] surveyed Computational Intelligence (CI) techniques to demonstrate the application of segmentation in the interdisciplinary research area. They discussed well-known CI techniques such as neural network, fuzzy logic, and genetic algorithms as well as they have also proved that CI-based approaches are cost-effective, time-saver, higher-efficient, and applicable in various engineering sectors [12, 13]. In 2021, researchers introduced an Internet of Things (IoT) device for automatic plant disease (galls) detection using the Fuzzy Based Function Network (FBFN) segmentation technique [14]. Similarly, a web-based tool was developed to identify mango leaf diseases such as Anthracnose using the RBF segmentation method. Thus, the segmentation techniques are universally implemented for different image-based detection techniques [15].

However, image segmentation is extensively used in human disease detection and diagnosis. For precise detection of the disease, initially, it requires identifying the region of interest from the captured images. In this study, a comprehensive description of the most important state-of-the-art medical image segmentation techniques is given. Here, we consider a mixed survey of segmentation for different types of medical data.

IMV-FCM, an enhanced multiscreen FCM clustering method, introduces a weighted adaptive learning technology to increase the flexibility of coordinating from diverse viewpoints. The algorithms might be able to learn from each view in an adaptive way that helps them better group brain tissue and deal with a noise like partial dimension distortions and grayscale that does not match up [16]. In 2018, Karegowda et al. [17] conducted research on the segmentation of brain tumor regions from MRI data. The authors conducted a comparative examination of FCM, Adaptive Regularized Kernel-based FCM approaches, PSO, and K-means and concluded that using PSO as swarm intelligence is a useful step. The results of the experiments showed that PSO-based segmentation is more accurate than FCM, Adaptive Regularized Kernel FCM, and K-means [17]. Arun Kumar et al. [18] created an improved automated approach for segmenting brain tumor regions and identifying them using K-means for the same objective. The goal of the authors was to improve the imaging enhancement at the pre-processing stage for precise brain tumor prediction [18]. Chander et al. [19] developed a framework for the segmentation of MRI images using K-means with Support Vector Machine (SVM) as the machine learning approach, and the overall accuracy was increased over earlier work [19]. In [20], the authors have discussed various methods such as traditional segmentation (Threshold, Fuzzy Theory, Region and edge detection), machine learning approach (KNN, Random Forest, SVM, Dictionary learning), and deep learning methods (CNN, FCM, Encoder/decoder). Although their analysis depicts that deep learning-based techniques such as FCM are superior to other traditional methods, they also include a supervised method that demands for manual labeling which required domain-specific knowledge [20].

In one of our previous studies, we have developed an IoT-based data collection system for skin lesions where we classified various skin lesions using deep learning-based ensemble algorithms [21]. For skin lesion segmentation, Yuan et al. [22] used the notion of Deep Fully Convolutional Neural Networks (DFCNN) with Jaccard distance. They employed the 19-DFCNN layer for self-training and the function of the new loss based on the Jaccard scale created by the researchers to re-measure using cross entropy to distinguish the lesion from the skin lesions. The findings of the studies imply that the upgraded classification approach outperforms conventional state algorithms, but that it requires more pre- and post-processing stages for greater accuracy [22]. Using CNN, Xie et al. [23] devised a reliable approach for extracting skin lesion bounds in the existence of distortions in digital images. Due to the use of a basic segmentation strategy, detecting the boundary of a skin lesion zone is slow, but this can be solved by utilizing a semantic segmentation technique [23]. In 2022, the authors trained a feature adaptive transformers network (FAT-Net) and managed to handle blurred boundary issues associated with lesions image. Yet FAT-Net may effectively extract local features and global true label whereas CNN are not capable of learning global true labels sufficiently [24]. Similarly, a neural network-based Multi-scale Residual Encoding and Decoding network (Ms RED) is used to handle blurred boundaries [25]. Thapar et al. [26] employed a segmentation framework using swarm intelligence with Grasshopper Optimization Algorithm (GOA) for feature extraction and successfully obtained 98.42% classification accuracy. Nevertheless, they only trained the model on three skin lesions images [26].

The existence of the nucleus in blood cells is used in determining Leukemia. In 2021, Daud et al. [27] used conventional algorithms such as watershed distance transform and Sobel edge detection algorithm for segmenting nuclei from microscopic images [27]. In another research, authors deployed a Global Local Entropy Histogram Equalization (GLEHE) based segmentation technique to identify Leukemia in blood cells [28]. Dhal et al. [29] provided a method for segmenting blood images for leukemia using the Stochastic Fractal Search (SFS) algorithm, which provides non-false positive segmented results. For image segmentation, the notion of K-means-based clustering is studied. The proposed scheme was compared to a previous clustering method, and the findings showed that the system’s performance was better in terms of efficiency, computational burden, and quality attributes [29].

Senthil Kumar et al. [30] used five algorithms to extract a plant region from very small lung images, including PSO, inertia-weighted PSO, guaranteed convergence PSO (GCPSO), K-means, and K-median. The flexible median filter outperformed the central filters, intermediate variables, and standard pre-processing stage, proving that it is best suited for medical CT imaging. In addition, employing the changing histogram balance improves the image brightness. Four algorithms are used to determine the quality of pre-processed images with improved quality. GCPSO has a high accuracy of 95.89 percent when visual results were confirmed with 20 lung sample images using MATLAB [30]. In 2021, van De Worp et al. [31] introduced deep learning-based two-step U-Net architecture for lung cancer segmentation from CT images. Although they performed the task only on 60 CT images [31]. The authors in [32] deployed 2-D Discrete Wavelet Transform (DWT) on the “LOTUS dataset” of lung tumor (CT images) and achieved a dice coefficient of 0.8472.

In this study, we are going to make it an intuitive and easy-to-understand framework for medical image segmentation.

3. Methodology

This section of the research article includes the procedural and working steps of the proposed model for Medical Image Segmentation using the Traditional Segmentation, SI, and CNN mechanisms. We focused on introducing a modified medical image segmentation approach using CNN as a deep learning and three distinct proposed architectures, which are as given in further sections.

3.1. Traditional Segmentation

In this phase, we evaluate the two clustering-based segmentation approaches such as FCM and K-means because it has many applications in medical research.

3.1.1. FCM-Based Segmentation

This scenario presents medical image segmentation using the concept of FCM as an unsupervised process. After segmentation, two parts of an image are formed known as the background and foreground part where the foreground is the ROI of any medical data such as MRI, microscopic, dermoscopic, CT-scan, etc. Here, we use some common pre-processing steps in the entire six scenarios of the proposed framework for comparative analysis, first is the color conversion using

$$ (1)

where G_Image is the grey level image that obtained after the color conversion from the color images (RGB ⟶ Red, Green, and Blue plane). After that, grey level mapping is initiated on the clipped region of the image for quality enhancement using the given equation:

$$ (2)

Equation (2) defines the average number of pixels in the medical image. Where X_{(region − xaxis)} is the total number of medical image pixels in a clipped region (X_CLIP). The clip limit (X_CL) of medical image enhancement is calculated using equation 3 then we apply the image enhancement of the further processing using the written Algorithm 1:

Algorithm 1: Enhancement of medical images.
•  

  Input: M ⟶ Medical-Image

•  

  Output: EM ⟶ Enhanced Medical Image

•  

  Start enhancement process

•  

  Set clip limit, X_CL = X_CLIP − X_AVERAGE

•  

   [Row, Col., and Plane] = Size (M)

•  

  If Plane > 1//Means image is color

•  

       CM = RGB to Grey (M)

•  

  Else

•  

  CM = M//No need of conversion

•  

  End-If

•  

  For I = 1 > X_CL

•  

  EM = Intensity (M (I), X_CL)

•  

  End-For

•  

  Return: EM as an enhanced medical image

•  

  End-Algorithm

After medical image enhancement in pre-processing phase, we move toward the segmentation using the FCM and the FCM algorithm written as below in Algorithm 2.

Algorithm 2: FCM-based Segmentation.
•  

  Input: EM ⟶ Enhanced Medical Image

•  

  Output: BM and FM (ROI) ⟶ Background and Foreground

•  

  Start FCM

•  

  Define cluster number (G = 2)

•  

  [Row, Col.] = Size (EM)

•  

  Segregate G = G1&G2 /Where G1 for BM and G2 for ROI

•  

  Rep = N//Define number of iterations for clustering

•  

  While Rep ≠ N (until max iteration not achieved)

•  

  For P = 1 ⟶ Row

•  

  For Q = 1 ⟶ Col

•  

  If EM (P, Q) = = G1

•  

  BM (P, Q) = EM (P, Q)

•  

  Else if EM (m, n) = = G2

•  

  ROI (P, Q) = EM (P, Q)

•  

  End-If

•  

  Adjust Centroid G using given Algorithm 2

•  

  G_mn =

•  

  Repeat and define FCM membership using given Algorithm 2

•  

  $$

•  

  End-For

•  

  End-For

•  

  End-While

•  

  Return: BM and ROI

•  

  End–Algorithm

The concept of FCM is dependent on the idea of consistently acquiring cluster centers by adjusting their positions using mean values that are given in equation 4 and allows clusters that are more flexible by introducing the possibility of partial memberships. The error function of FCM is written in the following equation:

$$ (3)

where fuzzy membership is denoted by μ_ij of X_i (image’s pixel) and the cluster identified by its center C_J, and here, k is a constant that defines the fuzziness of the resulting partitions.

$$ (4)

The steps involved in the algorithm of FCM-based segmentation are

•  

  FCM 1: recruit Ci as the cluster centers and Iteration N = 0

•  

  FCM 2: call FCM membership functions μ_ij according to equation 7

•  

  FCM 3: let N = N + 1 and assign new C_i as new centers

•  

  FCM 4: until the best convergence is not found, repeat steps 2 to 3.

Using this algorithm, we segment the ROI from the medical images and after segmentation of medical images; the obtained segmented result with original images shown in Figure 3.

3.1.2. K-Means Based Segmentation

This is the second scenario and we used K-means as a segmentation technique instead of FCM because K-means helps to provide better segmentation results as compared to the FCM that is shown in Figure 4. By utilizing the concept of K-means as a medical image segmentation technique, appropriate ROI from the medical images could be segmented but also K-means faced mix-up issues, and the algorithm of K-means is written in Algorithm 3.

Algorithm 3: K-means based Segmentation.
•  

  Input: EM ⟶ Enhanced Medical Image

•  

  Output: BM and FM (ROI) ⟶ Background and Foreground

•  

  Start K-means

•  

  Define cluster number (G = 2)

•  

  [Row, Col.] = Size (EM)

•  

  Segregate G = G1&G2//Where G1 for BM and G2 for ROI

•  

  Rep = N//Define number of iterations for clustering 

•  

  While Rep ≠ N (until max iteration not achieved)

•  

  For P = 1 ⟶ Row

•  

   For Q = 1 ⟶ Col

•  

    If EM (P, Q) = = G1

•  

      BM (P, Q) = EM (P, Q)

•  

     Else if EM (m, n) = = G2

•  

      ROI (P, Q) = EM (P, Q)

•  

    End-If

•  

  Adjust Centroid C using their mean

•  

  G = Average (BM, ROI) using the given Algorithm 3

•  

  $$

•  

  End-For

•  

  End-For

•  

  End-While

•  

  Return: BM and ROI

•  

  End-Algorithm

Based on the above written K-means algorithm in the ASBT system, we obtained better-segmented result as compared to the FCM-based ASBT system, and the results with the original MRI image are shown in Figure 4.

3.2. SI-Based Segmentation

In this scenario, we the concept of PSO as a SI approach because it is the most well-known optimization technique that helps to optimize the pixel-mixing problem faced by the FCM. Here, we present two different hybrid mechanisms named FCM and K-means with PSO for medical image segmentation.

3.2.1. FCM with PSO-Based Segmentation

In this scenario, we utilize the concept of PSO along with the FCM as a medical image hybrid segmentation. PSO is a powerful meta-heuristic technique that is favored by birds, mammals, and other insects that live or move in swarms. The best example of PSO inspiration is a flock of birds or a school of tiny fish that helps to reduce the pixel-mixing problem faced by the FCM during medical image segmentation by having very similar neighborhood pixel values. This hybrid segmentation method is based on the participation of individuals of each particle from many fields that are participating in the searching mechanism of a threshold value to minimize the mixed pixels. To solve a medical image segmentation problem, each particle modifies its threshold value based on its own and its neighbors’ experiences. Formally, each PSO particle P_I has a position P_I (t) at the time t instances in the search space, which change at time t + 1 by a velocity V_I(t). In the PSO algorithm, V_I(t) velocity is influenced by the best position V_BEST(t) visited by itself and P_ALL (t) the best position visited by all particles (we termed it “global best”). Each particle’s position is determined by a unique fitness function (Fit (Fun)), which is dependent on the segmentation issue and space dimension D.

$$ (5)

Kennedy and EberhaVrt, (1995) [9] established the PSO algorithm as an evolutionary image segmentation technique, and the algorithm of FCM with PSO segmentation is written below in Algorithm 4.

Algorithm 4: FCM with PSO based Segmentation.
•  

  Input: EM ⟶ Enhanced Medical Image

•  

  Output: BM and FM (ROI) ⟶ Background and Foreground

•  

  Start Hybridisation

•  

  Define cluster number (G = 2)

•  

  [Row, Col.] = Size (EM)

•  

  Segregate G = G1&G2//Where G1 for BM and G2 for ROI

•  

  Rep = N//Define number of iterations for clustering

•  

  While Rep ≠ N (until max iteration not achieved)

•  

  For P = 1 ⟶ Row

•  

   For Q = 1 ⟶ Col

•  

    If EM (P, Q) = = G1

•  

       BM (P, Q) = EM (P, Q)

•  

     Else if EM (m, n) = = G2

•  

       ROI (P, Q) = EM (P, Q)

•  

     End-If

•  

       Adjust Centroid G using given Algorithms 1 and 2

•  

    End-For

•  

   End-For

•  

  End-While

•  

  To optimize the ROI, here we used PSO algorithm and then initialize using the following parameters such as:

• (i)

  Iterations (T)

• (ii)

  Population-size (S)

• (iii)

  Lower-Limit (LB)

• (iv)

  Upper-Limit (UB)

• (v)

  Fitness function

• (vi)

  Number of selection (N)

•  

  Calculate size in terms of T = Row × Col.

•  

  Fitness function for the parameter optimization of PSO:

•  

  $$

•  

  For l = 1 ⟶ T

•  

  fs = EMRI(l)

•  

  $$

•  

  fit(fun) = using Algorithm 4

•  

  Threshold_value = PSO(P, T, LB, UB, N, fit(fun))

•  

  End-For

•  

  Define optimization iterations, O-Rep = N

•  

  While O-Rep ≠ N (until max iteration not achieved)

•  

  Thr = Threshold_value

•  

  Mask = Binary (ROI, Thr)

•  

   ROI Boundaries = Boundary (Mask)

•  

  For k = 1 ⟶ D

•  

   ROI = EM × ROI Boundaries

•  

  End-For

•  

  End-While

•  

  Return: BM and ROI as an improved background and foreground

•  

  End-Algorithm

Based on the above-written hybrid segmentation algorithm using FCM with PSO, we obtained better-segmented results as compared to the only FCM as well as K-means also and results with the original medical image shown in Figure 5.

3.2.2. K-Means with PSO-Based Segmentation

The concept of PSO along with the K-means clustering algorithm used as a medical image hybrid segmentation and the algorithm of K-means with PSO segmentation is written below in Algorithm 5.

Algorithm 5: K-means with PSO based Segmentation.
•  

  Input: EM ⟶ Enhanced Medical Image

•  

  Output: BM and FM (ROI) ⟶ Background and Foreground

•  

  Start Hybridisation

•  

  Define cluster number (G = 2)

•  

  [Row, Col.] = Size (EM)

•  

  Segregate G = G1&G2//Where G1 for BM and G2 for ROI

•  

  Rep = N//Define number of iterations for clustering

•  

  While Rep ≠ N (until max iteration not achieved)

•  

   For P = 1 ⟶ Row

•  

    For Q = 1 ⟶ Col

•  

     If EM (P, Q) = = G1

•  

      BM (P, Q) = EM (P, Q)

•  

     Else if EM (m, n) = = G2

•  

      ROI (P, Q) = EM (P, Q)

•  

     End-If

•  

   Adjust Centroid C using their mean

•  

  G = Average (BM, ROI) using the given Algorithm 5

•  

  $$

•  

       End-For

•  

  End-For

•  

  End-While

•  

  To optimize the ROI, here we used PSO algorithm and then initialize using following parameters such as:

• (i)

  Iterations (T)

• (ii)

  Population-size (S)

• (iii)

  Lower-Limit (LB)

• (iv)

  Upper-Limit (UB)

• (v)

  Fitness function

• (vi)

  Number of selection (N)

•  

  Calculate size in terms of T = Row × Col.

•  

  Fitness function using Algorithm 4 for the parameter optimization of PSO

•  

  For l = 1 ⟶ T

•  

  fs = EMRI(l)

•  

  $$

•  

  fit(fun) = using Algorithm 4

•  

  Threshold_value = PSO(P, T, LB, UB, N, fit(fun))

•  

  End-For

•  

  Define optimization iterations, O-Rep = N

•  

  While O-Rep ≠ N (until max iteration not achieved)

•  

  Thr = Threshold_value

•  

  Mask = Binary (ROI, Thr)

•  

  ROI Boundaries = Boundary (Mask)

•  

  For k = 1 ⟶ D

•  

   ROI = EM × ROI Boundaries

•  

  End-For

•  

  End-While

•  

  Return: BM and ROI as an improved background and foreground of medical image

•  

  End-Algorithm

Based on the above-written hybrid segmentation algorithm using K-means with PSO, we obtained better-segmented results as compared to the FCM, K-means, and FCM with PSO also and results with the original medical image shown in Figure 6.

3.3. CNN-Based Segmentation

This is the third module of implementation where we used again two different scenarios that are described below.

3.3.1. FCM with CNN-Based Segmentation

In this scenario, we utilize the concept of CNN as a deep learning mechanism along with the FCM as a medical image hybrid segmentation. This hybrid method is currently used in most of the existing medical image segmentation research. First, we train the model using lots of already segmented images in terms of background and foreground images having 3 dimensions (RGB). Usually, segmented medical images, which are fed into the neural network, are reduced in data dimensions, reduce the system processing time as well as complexity and help to reduce the over-fitting problems and hybrid CNN mechanism shown in Figure 7.

Figure 7 illustrated the process of medical image segmentation using the hybridization of FCM with CNN and the algorithm of K-means with PSO segmentation written below in Algorithm 6.

Algorithm 6: FCM with CNN based Segmentation.
•  

  Input: EM ⟶ Enhanced Medical Image

•  

  Output: BM and FM (ROI) ⟶ Background and Foreground

•  

  Start Hybridisation

•  

  Apply FCM and create BM and ROI

•  

  To optimize the ROI, here we used CNN and firstly we need to initialize using Epochs, Iterations, and Neurons (N) etc.

•  

  For I = 1 ⟶ T = Row × Col.

•  

  If EM (I) belongs to BM

•  

   Group (1) = Background

•  

  Else//EM belongs to ROI

•  

   Group (2) = Foreground

•  

  End-If

•  

  End-For

•  

  Call pattern net CNN for training of system using BM and ROI data

•  

  FCM-CNN = Pattern-net (Neurons)

•  

  FCM-CNN = Train (FCM-CNN, EM, Group)

•  

  ROI = Sim (FCM-CNN, ROI data)

•  

  If ROI (Pixel) = ROI

•  

   ROI = EM

•  

  Else

•  

   BM = EM

•  

  End-If

•  

  Return: BM and ROI as an improved segmented background and foreground of medical image

•  

  End–Algorithm

We obtained better-segmented results for the proposed hybrid mechanism of FCM with CNN as compared to the FCM, K-means, and improvisation in FCM with PSO, improvisation in K-means with PSO and results with the original medical image shown in Figure 8.

3.3.2. K-Means with CNN-Based Segmentation

The concept of CNN as a deep learning mechanism along with the K-means used in this scenario and similar to algorithm 6, a hybrid CNN mechanism with K-means is shown in Figure 9.

Figure 9 illustrated the process of medical image segmentation using hybridization of FCM with CNN and the algorithm of K-means with PSO segmentation written below in Algorithm 7.

Algorithm 7: K-means with CNN based Segmentation.
•  

  Input: EM ⟶ Enhanced Medical Image

•  

  Output: BM and FM (ROI) ⟶ Background and Foreground

•  

  Start Hybridisation

•  

  Apply K-means and create BM and ROI

•  

  To optimize the ROI, here we used CNN, and first, we need to initialize using Epochs, Iterations, Neurons (N) etc.

•  

  For I = 1 ⟶ T = Row × Col.

•  

  If EM (I) belongs to BM

•  

   Group (1) = Background

•  

  Else//EM belongs to ROI

•  

   Group (2) = Foreground

•  

  End-If

•  

  End-For

•  

  Call pattern net CNN for training of system using BM and ROI data

•  

  FCM-CNN = Pattern-net (Neurons)

•  

  FCM-CNN = Train (FCM-CNN, EM, Group)

•  

  ROI = Sim (FCM-CNN, ROI data)

•  

  If ROI (Pixel) = ROI

•  

      ROI = EM

•  

  Else

•  

      BM = EM

•  

  End-If

•  

  Return: BM and ROI as an improved segmented background and foreground of medical image

•  

  End–Algorithm

In comparison to FCM, K-means, FCM with PSO, CNN, and K-means with PSO, we achieved better-segmented results for the proposed hybrid mechanism of K-means with CNN, and results with the original medical image are displayed in Figure 10.

Finally, performance parameters for different types of datasets are calculated and compared using a comparison framework simulation in terms of Precision, Recall, F-measure, Accuracy, Error, MCC, DC, JC, and time.

3.4. Collected Dataset

3.4.1. Brain Tumor Segmentation (BraTS) Dataset

The sample images of the BraTS dataset are shown in Figure 11, it is a standard dataset obtained from “https://www.med.upenn.edu/sbia/brats2018/data.html” having MRI images [35]. For the simulation of the model, 50 DICOM files were converted into JPG format that is representing multi-frame superimposed brain images.

3.4.2. Acute Lymphoblastic Leukaemia Image Database (ALL-IDB) Dataset

The dataset contains 2008 images that were collected in September 2005 in the Image Processing Department of Computer Science-UniversitàdegliStudi di Milano” [36]. The ALL-IDB dataset of microscopic images is freely available for scientific research purposes from “https://homes.di.unimi.it/scotti/all/” and the sample of ALL-IDB dataset images is shown in Figure 12.

The used dataset contains approximately 39,000 blood counts, and oncologists labeled the lymphocytes. We resize the original microscopic images of blood samples into a size of 256 × 256 and a total of 2000 images were used in this research work.

3.4.3. ISIC-2018 Dataset

It contains the human lesion analysis toward melanoma detection and the dataset is in the form of dermoscopic images. The dataset is available from https://challenge2018.isic-archive.com/task1/training/ [37]. To capture images, the dermoscopic process is used which is an imaging technique to eliminate the surface reflection of human skin. It provides improved diagnostic accuracy and the sample of the ISIC-2018 dataset is shown in Figure 13.

3.4.4. CT-Scan Dataset

The database currently consists of an image set of 50 lung CT scans for research purposes which is publicly available from “http://www.via.cornell.edu/lungdb.html” [38]. A sample of dataset CT-scan images is shown in Figure 14.

After executing the methodology’s outlined steps, the performance has been evaluated in terms of several parameters, as discussed in the result and discussion section.

4. Results and Discussion

In this research work, we proposed a comparative framework for the medical image segmentation from various types of images such as MRI, Dermoscopic, Microscopic, and CT-scan images using the six different scenarios such as FCM, K-means, and improvisation of FCM using PSO, improvisation of K-means with PSO, improvisation of FCM with CNN and improvisation of K-means with CNN. Simulation results of the offered scenario are shown in Table 1 based on the quantities parameters.

There are five separate sets of data that included a varied number of images considering 50, 100, 500, 1000, and 2000 images. We have considered an equal number of images from each dataset to produce five specified subsets. Then we applied various segmentation techniques to the subsets of data. The concluded result has been described using different metrics such as precision, recall, accuracy, and F-measure, as shown in Table 1 and Figure 15.

From Table 1 and Figure 15, we observed that the simulation results of proposed frameworks, and hybridization of the K-means with CNN is superior to other modules in terms of the quantities parameters. Improvements in quantities parameters are clearly visible in Figure 15 and average accuracy is 85.72%, 86.06%, 87.54%, 88.64%, 92.21%, and 96.45% for FCM, K-means, FCM with PSO, K-means with PSO, FCM with CNN, and K-means with CNN, respectively. So, we can say that the effect of CNN on K-means for medical image segmentation is far better than other combinations. However, we need to validate the model based on similar parameters such as MCC, DC, JC, and computational time.

Therefore, the simulation results based on the similar values have been given in Table 2. In terms of similarities parameters, also K-means with CNN is superior to other modules for all the similarities metrics such as MCC, JC, and CD. The required time for K-means with CNN is slightly higher than other models. However, the change is extremely minute and may safely be ignored as a result. FCM with CNN is the second most successful segmentation technique based on both quantitative and similarity metrics. Also, it is transparent that the CNN-based optimized segmentation techniques performed better than both swarm intelligence and traditional methods.

Achieving maximum accuracy is the goal of the proposed framework with a fast response to segmentation with pre-processing. The proposed framework offers an extremely self-configurable and standalone mechanism with lots of deep learning interfaces. In addition, our suggested framework generated robust models for segmenting the region of interest in various issues. So, it is a universal framework for medical image segmentation. However, to validate the efficiency of the system, we need to compare it with state-of-the-artwork based on their accuracy in Table 3.

It is clear that the previous study was conducted on a single issue such as skin lesions, brain tumors, lung cancer, or leukemia (as shown in Table 3). However, our study aims to build a universal model for dealing with all the issues. In our case, we have found an average accuracy of 96.45 with K-means with CNN.

5. Conclusion

In this study, we have introduced a comparative framework for medical image segmentation with traditional, swarm intelligence and convolutional neural networks as a deep learning mechanism. This framework helps to design a real-time universal medical diagnosis system for various types of images such as MRI for brain data analysis, dermoscopic for skin, microscopic for blood leukemia, and CT-scan images for lungs. Here, we present a comparative study using six different scenarios such as FCM, K-means, improvisation of FCM using PSO, improvisation of K-means with PSO, improvisation of FCM with CNN, and improvisation of K-means with CNN. We proved the functionality of K-means with CNN is a powerful hybrid mechanism that achieves 96.45% accuracy whereas, other mechanisms achieve 85.72%, 86.61%, 87.54%, 88.64%, and 92.21% FCM, K-means, and improvisation of FCM using PSO, improvisation of K-means with PSO, improvisation of FCM with CNN and improvisation of K-means with CNN respectively. We also found that CNN-based optimized algorithms performed well compared to optimized swarm intelligence and/or traditional methods. We expect that in the future, it will aid in the transition of medical image segmentation from research laboratories to operational or real-time applications.

Conflicts of Interest

The authors declare that they have no conflicts of interest.