Minimally parametrized segmentation framework with dual metaheuristic optimisation algorithms and FCM for detection of anomalies in MR brain images

Highlights

• A novel combinational framework is introduced for the anomaly detection MRI.

• FCM, GLCM, ABC and Jaya algorithm combination enhances the robustness of segmentation output.

• GLCM features were used to enhance the ABC algorithm’s exploitation capability.

• JAYA algorithm with two improvements to its standard operations is used for the re-evaluation.

• Novel Jaya algorithm requires very few prerequisites to perform optimization.

• The proposed framework shows the good segmentation efficacy compared to its peer methodologies.

Abstract

Background

Early prognosis of a brain tumour may offer better life expectancy. Magnetic Resonance Imaging (MRI) coupled with an efficient machine learning segmentation technique has proven to be a reliable way of assessing tumours. In addition to the segmentation, the image is needed to optimise to achieve the desired results. In many cases, single-stage optimisation could not complete the search target owing to algorithm-specific limitations. To overcome this hindrance, the dual metaheuristic optimisation technique is widely used to detect tumour affected tissues.

Aim

This research emphasises brain tumour region detection using Fuzzy C-Means (FCM) clustering techniques and the segmented output enhancement using two different optimisation techniques, namely, Artificial Bee Colony (ABC) and the JAYA algorithm.

Methods

This methodology first deploys the FCM clustering technique to segment the tumour region in the MRI. Then, the initial stage of optimisation is done using the ABC algorithm with the help of texture features extracted from the segmented image through the Gray Level Co-occurrence Matrix (GLCM) technique. Lastly, a novel JAYA algorithm is deployed for the second stage of optimisation to provide precise segmentation with the support of global and local best solutions.

Result and Conclusion: The proposed framework delivers high accuracy in tumour detection. Besides, it has been proven by renowned evaluation metrics, such as Tanimoto Coefficient Index, and Dice Coefficient Index, which are up to 70.12% and 82.56%, respectively, competing with the contemporary methods used for the evaluations of MR brain images.

Introduction

Many radiologists say it's very challenging to diagnose brain-related problems; particularly tumour infiltration. Owing to the complex structure of the brain, it is also hard to distinguish the tumour-affected region and the edema region that surrounds it. Hence, oncologists and radiologists need an advanced automated diagnosing system to improve image visualisation for precise anomaly detection and infectious tissue fragmentation. Clinical experts prefer Magnetic Resonance Imaging (MRI) for diagnosis because it provides more information than other imaging modalities such as Computed Tomography (CT), X-Ray Imaging, Positron Emission Tomography (PET), and Ultrasound. However, MR Images may get distorted due to the movement of the patient during the scanning process and this challenge can be mitigated with the usage of a special sensing device mentioned in [49].

Manual segmentation takes a long processing time and can produce inconsistent findings due to intra-observer and inter-observer variability. Using a fully automated segmentation methodology with appropriate graphical tools, it is possible to segment brain lesions and other tissue regions into numerous separate parts without the need for human interaction. Many automatic segmentation [1], [37] algorithms have been developed in previous studies to comprehend tumour characteristics, like volume, size, shape, type and location. However, selecting an appropriate segmentation algorithm is very important to obtain the desired results. Although many optimisation algorithms have been developed in the literature, most of them are biased by the algorithm-specific parameters and may lead to inferior segmentation. To overcome the abovementioned problems, two minimally parameterized optimisation algorithms, namely, the Artificial Bee Colony and the novel JAYA algorithm, have been recommended in the study. The first stage optimisation enhances the segmented region based on the texture feature from GLCM, and the second stage optimises the output with a suitable modification in population size.

In general, five basic techniques are developed based on predefined attributes such as Edge, Region Growing, Threshold, Neural Network, and Fuzzy Logic for the fully automated image segmentation process. In the edge-based methods [2], Gradient (Sobel, Canny), Gaussian (Laplacian) based operators are used as edge-filter for the detection of dissimilarity in an image. In this method, pixels are categorized into the edge (converging to one) or non-edge (converging to zero), and thereby a target boundary is constructed. The accuracy level of this method depletes with the increase in homogeneousness (similarity in grayscale level distribution) between CSF, WM, and GM. Consequently, region-growing methods [3] have also been preferred for the tumour segmentation process in which a feasible solution is obtained for inhomogeneous brain tissues. In this method, if the nearby pixels obey the predefined intensity level guidelines, then the pixel is included in the region, and the procedure continues till all the pixels in the image are checked for similarity. The key limitation of the region-growing method is its high dependency on the order of scanning pixels. Different scanning positions may offer different segmentation outputs. In the threshold-based methods [4], the tumour region is isolated from the background by setting an appropriate threshold limit, which directly influences the sensitivity and selectivity property of the segmentation. In literature, various thresholding techniques were employed for image segmentation such as Otsu [5], maximum correlation, and optimal thresholding. Even a small variation in the threshold limit may impact the segmentation output adversely. Active Contour [5] model is considered widely for image segmentation due to its ability to delineate the region of interest from the background. However, this method needs a long duration to track the entire path from initial contour points.

In recent times, most researchers have preferred Artificial Neural Network (ANN) [6] based segmentation for producing robust results with less computational time at a moderate cost, backed by its different set of training parameters. ANN-based methods ensure higher efficacy when many slices are employed for the segmentation process. Though ANN has many merits in terms of multitasking capability, operational speed, and accuracy, it also has its demerits like unknown operational behaviour and selection of appropriate network structure. Many toolbox-based courseware were developed in the literature to ease the segregation of data. In [51], the authors developed the ADDIE methodology using a toolbox for the right identification and isolation of COVID-19 patients among the hajj pilgrims. Minimization of the number of instructions fed to the toolbox instructor can be achieved to abstain from human-driven error inputs. The fuzzy logic [7] is frequently used in image segmentation process due to its easy accessibility to large datasets and its ability to segment complex tissue structures. FCM clustering is utilized in this proposed methodology due to its salient features: (i) Membership value is assigned to data elements based on the fuzzy logic set, (ii) Allows data element to belong to multiple clusters if it satisfies the criterion of more than one clusters, (iii) Minimize the objective function in fewer iteration. Fuzzy-based clustering algorithms [8] are the best segmentation tool when there is a lot of vagueness and complex structures in the MR brain image. In literature, many FCM based (derived/modified) methodologies were developed such as Modified Kernel FCM [9], Fast Spatial FCM (FSFCM) [10], Type-2 adaptive weighted spatial FCM (AWSFCM) [11], Bat Algorithm with Fuzzy C-Ordered Means (BAFCOM) [12], K-means and modified fuzzy C-means [13] etc.

Feature extraction [50] is the method of grouping data points with the same attributes as frequency, pixel intensity, texture, and colour to reduce the dimensionality of the input test image for further processing. Choosing the appropriate traits from the test image is significant because the incorrect selection of attributes may degrade the overall segmentation results. In the proposed algorithm, the GLCM was utilized [14], [15] due to its effectiveness in redefining the image with reduced dimension (set of texture features) using its gray-level intensity.

Numerous optimisation techniques were deployed in the literature to enhance the segmentation accuracy, such as Evolutionary Algorithms [16], [17] and Swarm Intelligence Algorithms [18], [19], [20], [21], [22].

Unlike other optimisation algorithms, ABC requires only a minimal number of algorithm-specific parameters. For instance, a standard Genetic Algorithm needs three major control parameters, such as crossover mutation, mutation rate, and generation gap, apart from the population size and maximum evaluation number. Particle Swarm Optimisation (PSO) algorithm also needs control parameters like cognitive, acceleration coefficient, inertia weight, and particle count apart from the maximum velocity. On the other hand, the ABC algorithm needs only one control parameter in addition to the two algorithm-specific parameters like the number of food sources and maximum search count. Secondly, the ABC algorithm is easy and simple to hybridize with other algorithms and the third advantage is its fast convergence offered. Owing to the above-mentioned salient features, the authors have chosen the Artificial Bee Colony [23] algorithm for the initial level of optimisation.

Though the ABC algorithm has merits like the ease of implementation and the ability to explore local solutions with a fast convergence rate, it also has its demerits in premature convergence during the search period. In many cases, single-stage optimisation could not complete the search target owing to algorithm-specific limitations, which can be resolved by developing a two-stage optimisation [34] to examine the brain abnormality in MR brain images. To overcome the above-said dissensions, the JAYA algorithm [24] was used as the second optimisation technique to enhance segmented output further. It has the following advantages, such as (i) parameter less tuning, (ii) taking very little iteration to minimize the objective function, and (iii) requires very less memory to process voluminous slices. In the study, a hybrid segmentation framework was developed by improvising the métiers and reducing the weaknesses influenced by each algorithm to detect tumour regions effectually.

It is well known that no optimisation technique is superior to the other(s) in solving all optimisation related problems. However, literature shows that an optimisation technique may address a specific problem effectively. On the back of this, an integrated metaheuristic algorithm, a combination of the improved ABC algorithm and the novel JAYA algorithm, was developed in the research. The motivations for this study are.

• To develop a fully automatic system with dual metaheuristic optimisation techniques to segment brain MRI with a higher success rate than the single optimisation algorithm,

• To address the exploitation and exploration limitations confronted by the optimisation algorithms during search operation with suitable modification to its standard operations, and

• To reduce the overall convergence computation cost by using the fewest possible control parameters.

To achieve the above-mentioned motivations, the contributions made to this research are as follows:

• To the best of knowledge, this is the first work which integrates two metaheuristic optimisation algorithms at different stages to enhance the segmentation output with minimal algorithm specific control parameters, as shown in Fig. 1.

• For the first time in literature, GLCM features were used to enhance the ABC algorithm’s exploitation capability during its search operation.

• First and foremost, in literature survey made, the novel JAYA algorithm with two improvements to its standard operations is introduced in the spatial domain for re-evaluation of the segmented image.

• Assessing the outcomes in both the normal and noise-embedded images to show the ability of the proposed methodology.

The paper's organization as follows, section 2 elaborates the material used and methodology of the recommended algorithm. In section 3, a detailed explanation of the proposed novel Dual Optimisation Algorithm (DOA) and their respective pseudocodes is discussed. Section 4 briefs results and compares the proposed algorithm's performance to other competing methods available in the literature. Lastly, the conclusion and future work planned are presented in section 5.