Image Anal Stereol 2018;37:249-275 doi: 105566/ias.1865 Original Article 249   NEW BACTERIA FORAGING AND PARTICLE SWARM HYBRID ALGORITHM FOR MEDICAL IMAGE COMPRESSION  G. VIMALA KUMARI,1, G. SASIBHUSHANA RAO2, B. PRABHAKARA RAO3 1Department of Electronics and Communication Engineering, M.V.G.R.College of Engineering, Vizianagaram- 535005, India; 2Andhra University College of Engineering, Visakhapatnam-530003, India; 3Jawaharlal Nehru Technological University, Kakinada, Kakinada-533003, Andhra Pradesh, India e-mail: vimalakumari7@gmail.com, sasigps@gmail.com, drbprjntuk@gmail.com  (Received November 25, 2017; revised May 29; 2018; accepted September 9, 2018) ABSTRACT  For perfect diagnosis of brain tumour, it is necessary to identify tumour affected regions in the brain in Magnetic Resonance Imaging (MRI) images effectively and compression of these images for transmission over a communication channel at high speed with better visual quality to the experts. An attempt has been made in this paper for identifying tumour regions with optimal thresholds which are optimized with the proposed Hybrid Bacteria Foraging Optimization Algorithm (BFOA) and Particle Swarm Optimization (PSO) named (HBFOA-PSO) by maximizing the Renyi’s entropy and Kapur’s entropy. BFOA may be trapped into local optimal problem and delay in execution time (convergence time) because of random chemotaxis steps in the procedure of algorithm and to get global solution, a theory of swarming is com- menced in the structure of HBFOA-PSO. Effectiveness of this HBFOA-PSO is evaluated on six different MRI images of brain with tumours and proved to be better in Peak Signal to Noise Ratio (PSNR), Mean Square Error (MSE) and Fitness Function. Keywords: Bacteria Foraging Optimization Algorithm (BFOA), Grey Wolf Optimization (GWO), Kapur’s entropy, Moth-flame Optimization (MFO) Algorithm, Particle Swarm Optimization (PSO), Renyi’s entropy. INTRODUCTION In this modern world human beings are very much busy with their personal and professional life and undisciplined food and sleeping timings and busy life style which are causing health disorders. Brain tumour is one such dangerous health disorder and it is due to growth of cells abnormally in the tissues of the brain and can directly destroy or damage brain cells by producing inflammation. Brain tumours are classified by their size and type of tissue involved. In the areas of human brain image analysis, recognition of tumour region and segmentation of tissue organization tend to be a demanding task. Computerized segmentation of Magnetic Resonance (MR) brain images would be of immense help to radiologists, as they reduce the difficulties developed due to human interface and offer quicker segmentation results. Computerized algorithms offer negligible time duration and slighter manual involvement to a radiologist during clinical diagnosis. In addition, huge volumes of patients’ information can be evaluated by computerization. For this purpose one algorithm called Hybrid Bacteria Foraging Optimization Algorithm and Particle Swarm Optimization (HBFOA– PSO) is proposed for effective and efficient image segmentation for identification of the brain tumour and segmented image is transmitted over a communication channel after successful compression and is received at the receiver section by a radiologist for diagnosis of brain tumour and for necessary action. This process is called telemedicine and its main objective is to provide clinical care from a distance. This benefits many rural areas and also at times of emergencies when doctor’s presence is essential. But further development of this technology is becoming slow due to the limited narrow transmission bandwidths. As the presence and development of such technologies are crucial, a lot of research is being carried out for further improvement in effective usage of such technology. So in this paper, an HBFOA–PSO based image thresholding by maximizing Renyi’s entropy and Kapur’s entropy is proposed for efficient and effective results of image thresholding in better image compression, which helps reduce the Bit rate for transmission as well as maintains an appreciable amount of quality or fidelity of the image. Image thresholding is a process of optimizing similar regions   VIMALA KUMARI G ET AL: Hybrid Algorithm for Medical Image Compression 250   in an image which results in effective clustering of image, hence better image compression is achieved by running cascaded runlength coding and arithmetic coding on cluster image. Performance of the runlength coding and arithmetic coding depends upon the number of similar clusters and probability occurrence of same cluster centroids in an image respectively. So, its performance depends upon effective clustering technique. MATERIAL AND METHODS CONTRIBUTION In this section a detailed description is given about the methods of MRI image tumour recognitions and methods to compress the identified regions at high bit rate with better reconstructed image quality. In first section, methods of MRI image segmentation are explained and in second section methods of image compression are explained. RELATED WORK IN TUMOUR IDENTIFICATION Stochastic threshold of MRI image for tumour identification is done by combining region based level sets globally and three established energies (uniform, separation, and histogram) in a local framework (Lubna et al., 2017). An automated brain tumour threshold model based on maximum a posteriori probabilistic (MAP) estimation and likelihood probability of the model is estimated by sparse coding and dictionary learning. The Markov random field (MRF) is introduced into the prior probability. The MAP is converted into a minimum energy optimization problem and graph cuts are used to find its solution (Yuhong et al., 2016). Irem et al., proposed a Principal Component Analysis (PCA) based K-means and fuzzy C-means (FCM) clustering for MRI image thresholding and results are compared with five PCA algorithms such as PCA, Probabilistic Principal Component Analysis (PPCA), Expectation Maximization Based PCA (EM-PCA), Generalize Hebbian Algorithm (GHA), and Adaptive Principal Component Extraction (APEX) and proved EM-PCA and PPCA results are effective with the two clustering algorithms (Irem et al., 2017). The brain tumours are identified and marked with a novel technique which is proposed by solmaz (Solmaz and Farshad, 2017). The 3D images are pre-processed with the help of cascaded histogram matching and bias field correction. After this the required areas are extracted from background. Local binary pattern and orientation gradients are used for learning. Potential Field Clustering (PFS) is one which is based on concept of potential field and evaluates the performance of the different methods on the brain tumour MRI benchmark database (Ivan and Iker, 2017). Statistical fusion based image thresholding is proposed for identification of abnormalities in MRI images by seed selection, region growing and image fusion. The proposed technique is tested for performance analysis on different data base and different tumour effected images (Badri et al., 2016). Nooshin and Miroslav proposed automatic tumour thresholding in single- spectral MRI using a texture-based and contour-based algorithm (Nooshin and Miroslav, 2017). Sudip et al., proposed a conditional spatial fuzzy C- means (csFCM) clustering algorithm for thresholding of MRI images by incorporating local and global spatial information into a weighted membership function and results are compared and proved better than K-means and FCM algorithms in terms of validity functions, threshold accuracy, tissue threshold accuracy, and receiver operating characteristic (ROC) curve (Sudip et al., 2015). Edge detection of X-ray images using Multiresolution Analysis (MRA) based biorthogonal wavelets is more preferable when compared with orthogonal wavelets because of more flexibility (GS Rao et al., 2016). Brain tumour is identified with efficient thresholding of MRI images by Bacteria Foraging Optimization (BFO) with modified fuzzy K- means algorithm (MFKM) and results are compared with Particle Swarm Optimization (PSO) based fuzzy C-means algorithm (PSO based FCM), MFKM and conventional FCM and proved better in terms of sensitivity, specificity, Jaccard Tanimoto Co-efficient index (TC) and Dice Overlap Index (DOI), computational time and memory requirement (Anitha et al., 2017). Computed Tomography (CT) images are classified by Support Vector Machine (SVM) with different kernel functions and Sequential Minimal Optimization (SMO) and segmented classification is further performed by the Modified Region Growing (MRG) with threshold optimization. These thresholds are optimized with Harmony Search (HS), Evolutionary Programming, (EP) and Grey Wolf Optimization (GWO). In terms of sensitivity, specificity, and accuracy GWO is better compared to others (Ramakrishnan and Sankaragomathi, 2017). MRI images are segmented with Teaching Learning Based Optimization (TLBO), entropy value, and level set/active contour and TLBO achieved better values in Jaccard index, dice co-efficient, precision, sensitivity, specificity, and accuracy (Rajinikanth et al., 2017). Crow Search Algorithm (CSA) based image thresholding by maximizing cross entropy for MRI image threshold and proved better with other techniques in terms of quality and consistency (Diego et al., 2017). Image Anal Stereol 2018;37:249-275  251   Sathya and Kayalvizhi took Kapur’s and Otsu’s entropies as an objective functions and which are optimized with Bacterial Foraging Optimization Algorithm (BFOA) for effective and efficient image thresholding (Sathya and Kayalvizhi, 2011). Further, in order to improve the convergence speed and global searching ability of BFOA, they modify the swarming step and reproduction step, thereby improving the robustness of bacterial foraging (BF) and achieved fast convergence. Same authors employed few modifications to BF for threshold of brain magnetic resonance images by adaptively varying the step size of bacteria instead of fixed step size followed by ordinary bacterial foraging (Sathya and Kayalvizhi, 2011). RELATED WORK IN IMAGE COMPRES- SION Image compression is achieved by appropriate image thresholding and these thresholds are obtained with a principal of moment preserving and was proposed by Chen and Wen (1998) (Chen-Kuei and Wen-Hsiang, 2015). The proposed method achieved a high compression ratio with better reconstructed image quality. An image compression method which consumes less time and follows a strategy where thresholds are optimized with optimization techniques for which objective function is distortion (Kaur et al., 2007). Birge–Massart thresholding is inbuilt thresholding technique which is used for image compression and obtained results are compared with the uni-modal thresholding in terms of reconstructed image quality and compression ratio (Siraj, 2015). In (Tahere and Mohammad, 2009) Electrocardiography (ECG) signals are compressed by transforming the signal with the help of discrete wavelet transform. Another kind of image compression where image to be compressed is transformed to frequency domain with the help of bandlet and required bandlet coefficients are obtained with type II Fuzzy thresholding and results are compared with the ordinary thresholding (Rajeswari, 2012). Prashant and Ioana proposed a non-uniform thresholding and observed the effects of thresholding on reconstructed image quality (Prashant and Ioana, 2003). Tony and Zhou proposed a technique for noise removal and image compression in wavelet domain thresholding which is based on Partial Differential Equation (PDE) and it takes the advantage of variations in framework (Tony and Zhou, 2007). Image compression can also be performed with Multistage Lattice Vector Quantization (MLVQ) and by thresholding of Discrete Wavelet Transform (DWT) coefficients. Proposed combination tries to minimize the quantization error and its computational complexity is less compared to ordinary VQ (Salleh and Soraghan, 2007). Kaveh et al., proposed a 2-D discrete wavelet transform based image thresholding by optimal thresholding the DWT coefficients with Particle Swarm Optimization (PSO) for image compression. They did three level decomposition of DWT and 62.5% of thresholds are assigned and optimized for the approximation coefficients and the remaining 37.5% equally assigned to horizontal, vertical, and diagonal coefficients (Kaveh et al., 2015). They compared the results with the Set Partition in Hierarchical Tree (SPHIT), Chrysafis, JPEG and JPEG-2000 and proved better in Peak Signal to Noise Ratio (PSNR) and Bits per Pixel (BPP). In this paper, HBFOA-PSO based brain MRI image thresholding is proposed for image compression by optimizing the Renyi's entropy and Kapur’s entropy for the first time and obtained results are compared with other optimization techniques such as BFOA, PSO, Moth-flame Optimization (MFO) Algorithm and GWO. Compressed image is further coded with runlength coding followed by arithmetic coding. Objective function value, standard deviation, Structural Similarity Index Measure (SSIM), PSNR, Weighted Peak Signal to Noise Ratio (WPSNR) and computational complexity are considered for the performance evaluation of proposed HBFOA-PSO based image thresholding. In all parameters the proposed algorithm performance is better as compared to other BFOA, PSO, MFO and GWO. PROBLEM FORMULATION OF OPTIMUM THRESHOLDING METHODS Initially all the required thresholds are selected randomly and these thresholds are optimized with the help of optimization techniques. In this paper, the objective function to be maximized with the optimization techniques is entropy techniques. The entropy techniques which are used in this paper are Renyi's entropy and Kapur’s entropy. After successful optimization of thresholds, image is partitioned into object and background. Assume gray scale image which contains L gray levels with range between 0 to L-1(0, 1, 2, . . . , (L - 1)). Then probability occurrence of pixel Pi = h(i)/N (0 1, optimal threshold *3t with Renyi's entropy is equal to threshold with the entropic correlation method. The optimal threshold value of the Renyi's entropy is calculated by the following formula with * 1t , * 2t , and * 3t . * [1] [1] 1 [2] 2 [3] [3] 3 1 1 1[ ( ) ] [1 ( ) ] 4 4 4c t t p t w t w t p t w        (7) Where t[1], t[2] and t[3] are order statistics of the gray values *1 ,t * 2t and * 3t , [3] [1] 1 ( ) , ( ) ( ) t i i p t p w p t p t     , and [1] [2] [2] [3] [1] [2] [2] [3] 1 2 3 [1] [2] [2] [3] [1] [2] [2] [3] (1,2,1) 5 5, (1,2,1) 5 5, ( , , ) (0,1,3) 5 5, (3,1,0) 5 5. if t t and t t if t t and t t if t t and t t if t t and t t                         (8) The optimal threshold value *ct can be viewed as an image dependent weighted average of *1t , * 2t and * 3t and thus * * * * * * * 1 2 3 1 2 3min{ , , } max{ , , }ct t t t t t t  that is, *[1] [3]ct t t  . This shows that the maximum entropy sum method or the entropic correlation method does not succeed in providing a good threshold value for a gray scale image but the Renyi's entropy provides a better threshold value. CONCEPT OF KAPUR’S ENTROPY Kapur developed an algorithm for bi-level thresh- olding which is as follows: The objective function is J(t) = H0 + H1 (9) Where 1 0 0 0 0 ln t i i i p pH w w     , 1 0 0 t i i w p    1 1 1 1 ln L i i i t p pH w w     , 1 1 L i i t w p    When objective function Eq. (9) is maximum then thresholds are optimal threshold. For multi-level thresholding Eq. (9) becomes J(t0, t1,……. tm) = H0 + H1 + H2 + …… + Hm (10) Where m is number of thresholds to be optimized and Image Anal Stereol 2018;37:249-275  253   1 1 0 0 0 0 ln t i i i p pH w w     , 1 1 0 0 t i i w p    2 1 1 1 1 1 ln t i i i t p pH w w     , 2 1 1 1 t i i t w p     3 2 1 2 2 2 ln t i i i t p pH w w     , 3 2 1 2 t i i t w p     1 ln m L i i m i t m m p pH w w     , 1 m L m i i t w p    Image compression with the Renyi’s entropy and Kapur entropy with two level thresholding proved efficient, but when threshold levels are increasing (multilevel thresholding) Renyi’s entropy and Kapur’s entropy takes much time for simulation and time increases exponential with levels. To improve the performance of Renyi’s entropy and Kapur’s entropy and to reduce the simulation time, few applications of soft computing techniques such as BFOA, PSO, MFO, and GWO for image thresholding are proposed, hence effective image compression. These techniques are to maximize the Renyi’s entropy and Kapur entropy as given in Eq. (5) and Eq. (9). PROPOSED HBFOA-PSO The main objective of this paper is to get optimal thresholds which leads to a better reconstructed image quality at high compression ratio. The optimal thresholds are obtained by cascaded combination of bacterial foraging optimization algorithm and particle swam optimization. The bacterial foraging optimization algorithm is global search algorithm but it may be trapped into local optimal problem and delay in execution time (convergence time) because of random chemotaxis steps in the procedure of algorithm and to get global solution, a theory of swarming is commenced in the structure of HBFOA-PSO. In order to find the advantage of proposed HBFOA-PSO approach, the results are compared with individual BFOA and PSO approaches. Each of algorithms is explained below. OVERVIEW OF BACTERIA FORAGING OPTIMIZATION ALGORITHM In 2002, a new optimization technique is proposed by Passino which is based on the foraging behavior of bacteria called Bacteria Foraging Optimization Technique (BFOA) (Passino, 2002). E.Coli bacteria always searches for nutrients to enhance the energy levels per unit time. Some of the bacteria search for nutrients by communicating with each other. In general, bacteria search for nutrients with the help of tumbling or swarming and chemotaxis step. The BFOA performance is better than other optimization techniques because of its advanced algorithm structure. In BFOA step of walk follows a Gaussian distribution function in searching food instead of Levy flight which is used in cuckoo search algorithm. In this paper the objective function which is optimized with BFOA is entropy (Renyi's or Kapur’s entropy). The bacteria moves in such a way that in each iteration the objective function is maximum. Because of this, each bacterium carries different objective function value in each iteration. Among all the objective function values of bacteria, the highest value is carried to next iteration. The remaining bacteria always try to move towards the highest objective function value bacteria and attain further highest values after successful final iteration. In this way all the bacteria attain global optimal solution. The BFOA attain this optimal solution with four cascaded steps: 1. Chemotaxis, 2. Swarming, 3. Reproduction and 4. Elimination-dispersal. The four steps in BFOA are explained below. 1. Chemotaxis: Chemotaxis step is a crucial step in BFOA while searching for food and it illustrates intelligence applied by the bacteria while searching for food. The bacteria try to move towards the better solution by taking either thumbling or swimming. In BFOA, each bacteria move to its better position by taking 8-neighbourhood positions derivative. After derivative, it finds which bacteria has maximum objective function and remaining bacteria follow the maximum objective function bacteria. The steps of Chemotaxis are as follow: Tumbling: In this step bacteria moves randomly in a particular direction where high nutrients are available in the search space. Initially all the bacteria are having natural nutrients. This process is known as tumbling and is shown in Fig. 1. Fig. 1. Tumbling of bacterium Swimming up: After successful and sufficient nutrient from tumbling, the bacteria move in the same direction if nutrients are further increasing or else it take Counter clockwise rotation TUMBLE   VIMALA KUMARI G ET AL: Hybrid Algorithm for Medical Image Compression 254   swimming step. This swimming movement is called swimming up. Swimming down: If the direction of movement decrease the bacteria nutrients then movement is called swimming down. When bacteria experience swimming down then immediately it changes its direction. This process is shown in Fig. 2. Fig. 2. Bacterial swim The chemotactic step of bacteria is mentioned in Eq. (11). θ j 1, k, l θ j, k, l C i . ∆ ∆ (11) Where C (i) is step size and Δ (i) is the random number lying between [0,1]. 2. Swarming: The collection of bacteria which are at higher nutrients will send information through signal to other bacteria. So rest of the bacteria will try to move towards the higher nutrients direction and avoids the direction of movement towards the lower nutrients. This step of process is called swarming and is described in below equation. 𝐽 𝜃, 𝑃 𝑗, 𝑘, 𝑙 ∑ 𝐽 𝜃, 𝜃 𝑃 𝑗, 𝑘, 𝑙 ∑ 𝑑 exp 𝑊 ∑ 𝜃 𝜃 ∑ ℎ exp 𝑊 ∑ 𝜃 𝜃 (12) 3. Reproduction: All the bacteria fitness values are shortlisted ascending or descending based on maximization problem or minimization problem respectively. In this paper, the objective function is to be maximized so, all the bacteria are arranged in ascending order based on their fitness values or objective function value. In this step the bacteria with lowest nutrients would die. In general, around half of bacteria die in this step and new bacteria is generated by asexual between two highest nutrients bacteria. Indirectly in this step among the available bacteria, half of bacteria die and these are replaced with newly generated bacteria to always maintain constant bacteria in search space. The process of new bacteria generation is called conjugation. 4. Elimination-dispersal: In some cases bacteria may experience a sudden change in environmental conditions like hike in temperature or in humidity. Then bacteria undergoes third step i.e. reproduction where bacteria may die because of sudden changes and new bacteria are generated by a sexual relation between two bacteria. Some bacteria may move to the nearest safest place. BFOA algorithm: For l = 1:Ned For k = 1:Nre For j = 1:Nc For i = 1:S J(i, j, k, l) = J(i, j, k, l)+Jcc[𝜃 𝑗, 𝑘, 𝑙 , 𝑃 𝑗, 𝑘, 𝑙 ∆ 𝑖 , 𝑚 1,2, … … … , 𝑝 𝜃 𝑗 1, 𝑘, 𝑙 𝜃 𝑗, 𝑘, 𝑙 𝐶 𝑖 . 𝛥 𝑖∆ 𝑖 ∆ 𝑖 Calculate J (i, j+1, k, l) m = 0 While m𝐽 Update 𝐽 𝜃 𝑗 1, 𝑘, 𝑙 𝜃 𝑗, 𝑘, 𝑙 𝐶 𝑖 . 𝛥 𝑖∆ 𝑖 ∆ 𝑖 else m = Ns End if End while End i-for End j-for End k-for For i =1:S 𝐽 =∑ 𝐽 𝑖, 𝑗, 𝑘, 𝑙 End i-for End k –for For i = 1:S SWIM Clockwise rotation Image Anal Stereol 2018;37:249-275  255   If rand () < Ped Eliminate bacterium and initialize randomly its replacement End if End i-for End l-for End BFOA OVERVIEW OF PARTICLE SWARM OPTIMIZATION PSO is proposed by Kennedy and Eberhart in the year 1995 and it is a stochastic approach under swarm intelligence that mimics how the particles are flying to get the best food location (Kennedy and Eberhart, 1995). Each individual particle adaptively updates their velocity and position within the search depends on the previous experience of its own search and the experiences of other particles in the population. Each particle is assigned with a memory by which it can store the best food location it ever visited during its journey. Its best food location is named as Pbest and the best food location of the group taken as one is stored as Gbest. The initial positions for Pbest and Gbest are different. It is proved to give the best results in obtaining the global minima or maxima. However, obtaining the global minima about the optimum value is a challenging issue, whenever multiple minima exist. This algorithm does not involve cross-over or mutation operators. It only depends on the initialization of the control parameters, the size of the swarm, the objective function and the maximum number of iteration. It does not depend on the initial conditions and the gradient values. The advantages of using PSO are computationally less expensive, much simple to implement, Less CPU time and memory requirement. The modified velocity of each particle is given by 𝑣 𝑡 1 𝑣 𝑡 𝑐 𝑟 𝑃𝑏𝑒𝑠𝑡 𝑥 𝑡 𝑐 𝑟 𝐺𝑏𝑒𝑠𝑡 𝑥 𝑡 (13) The modified position of each particle is given by 𝑥 𝑡 1 𝑥 𝑡 𝑣 𝑡 1 (14) PSO Algorithm: Step 1: Initialization of each individual particle in the population with random position and random velocity. Step 2: Calculate the objective function of each particle with Eq. (5) and Eq. (9). If the current cost is higher than the best value so far calculated, then it is stored in Pbest. Step 3: Choose the particle with the highest objective function value of all particles. The position of this particle is Gbest. Step 4: Calculate the new velocity and position of each particle according to the Eq. (13) and Eq. (14). Step 5: Repeat the above steps from 2-4 until maximum iterations or maximum criteria is not attained. HBFOA–PSO ALGORITHM The HBFOA–PSO algorithm combines BFOA and PSO algorithms, so it takes the advantages and disadvantages of both techniques. The aim is to share information between PSO and BFOA that leads to generation of healthy bacteria by dispersal and elimination. The major drawback with the BFOA is, step of tumbling is random so achieving a global solution is difficult. Where as in the proposed hybrid BFOA-PSO, the step of tumbling is not random and these tumbling steps are optimized with the PSO. The global best solution or better suitable tumbling step from the PSO is given as input to BFOA. Tumbling step is updated when BFOA is in first step. The parameters required for hybrid BFOA-PSO are given below. Step 1: Initialization of parameters for both BFOA and PSO: p = Dimensions of the problem; S = population size or number of particle in case of PSO and number of bacteria in case of BFOA; Ns = swimming length after tumbling operation when bacteria is in chemotaxis loop; Nc = stopping criteria or maximum iterations of the algorithm; Nre = Maximum number of reproduction steps; Ned = Maximum number of steps in elimination and dispersal loop; Ped = the probability of elimination and dispersal; C(i) =step of walk in tumbling stage and is random in BFOA algorithm; 𝑑 , 𝑤 , ℎ , 𝑤 = Bacteria attractive and repellent coefficients; Δ (p, i) = Bacteria direction in current iteration; P (i, j) = Bacteria position in current iteration;     𝑐 , acceleration r1, r2 = Ran PSO; Step 2: Elim Step 3: Repr Step 4: Chem Substep a: F step which i Calculate all Then new fit J(i, j, k, l) = Assign J_las Substep take either tu generated nu the bacteria onwards tum with PSO. B 𝜃 𝑗 1, 𝑘, Where j, k reproduction Which leads in tumbling Calculate J ( Jcc[𝜃 𝑗, 𝑘, 𝑙 Swimming s i. Assume m ii. While m Let m= If J(i, j+ and 𝜃 𝑗 Use this 𝜃 (j + iii. Else, le statement fo Substep go to substep Step 5: Ev bacterium an 𝑐 = PSO coefficient r dom number ination and d oduction loo otaxis loop: or i = 1, 2, s as follows bacteria obj ness function J(i, j, k, l) + t = J (i, j, k, l b: For i = 1 mbling or sw mber betwee or for all pop bling positi acteria move 𝑙 𝜃 𝑗, 𝑘, and l are & Eliminati ith bacteria w stage i, j, k, l) = J( , 𝑃 𝑗, 𝑘, 𝑙 tage = 0 (counte < Ns m + 1 1, k, l) >𝐽 1, 𝑘, 𝑙 new 1, k, l) to com t m = Ns. r swim. c: Go to nex b to proces aluate the d global bes VIMA cognitive co espectively; s between 0 t ispersal loop p: k = k + 1. j = j + 1. . . . , S, ith b ective value, Jcc (𝜃 𝑗, 𝑘, 𝑙 ). , 2, . . . , S b imming Δ(i n 0 to 1 in f ulations. Fro on and direc s towards be 𝑙 𝐶 𝑖 . ∆ index of on and dispe ill move w i, j, k, l)+ r for swim le , let 𝐽 = | J 𝜃 𝑗, 𝑘, 𝑙 pute new J ( This is the t bacterium ( s the next bac local best p t position. LA KUMARI nstant and o 1 in case of : l = l + 1. acteria move J(i, j, k, l); , 𝑃 𝑗, 𝑘, 𝑙 ); acteria decid ), that is rand irst iteration m second ite tion are opti tter direction ∆ Chemotaxis rsal respectiv ith a step siz ngth). (i, j+1, k, l)| 𝐶 𝑖 . ∆ ∆ i, j + 1, k, l). end of the i + 1) if i S terium) osition for G ET AL: Hyb 256 social with ed to omly for all ration mized with step, ely. e C(i) while (i.e., each St us St un St 2. ∑ nu w ch va 𝑗 be pl St St w ba Fi Co PN siz A “M re ar eq th 1. te A di ag rid Algorithm ep 6: Updat ing PSO. Up ep 7: If j < N til life of the ep 8: Reprod Substep a: . . S, find 𝑗 𝑖, 𝑗, 𝑘, trients it go as at avoidin emotactic pa lue (Lower 𝑗 Substep b values are st 𝑗 value aced at the sa ep 9: If k < N ep 10: Elimi ith probabil cterium. g. 3: Test ronary T1 A ET f) Menin RESUL For evalua e 256x256; strocytoma”, eningioma’ solution are e captured uipment, in ickness 1mm 5 Tesla ma chnology us vanto-Tim te agnosis of br e 3, 32, 35 for Medical I e the velocit date the vect c, go to step bacteria is o uction: For the give health of 𝑙 . 𝑗 give t over its lif g noxious su rameters C value me : The Sr = S removed an split. New me location re, go to step nation and d ity 𝑃 , eli images (cloc strocytoma gioma TS tion of expe namely “A “Glioma”, ’ with a 8 taken as sho from Sie which imag ×1mm×1mm gnetic field ed in this chnology. Ex ain tumour o , and 42. T mage Compr y and positio or ∆(p, i). 4 and cont ver. n k and l, an bacterium i, s a measure etime and ho bstances. S (i) in order o ans lower he /2, bacteria d other Sr b bacteria tha as their paren 3. ispersal: Fo minate and kwise) a) c) Glioma d riments, six strocytoma”, “Metastatic” bits per p wn in Fig. 3 mens-Area e is captur by 48 mul intensity. equipment periments ar n four differ he program ession n of bacteri inue chemota d for each i = 𝑗 as 𝑗 of how m w successfu ort bacteria f ascending alth). with the low acteria with t are made t. r i = 1, 2. . disperse e Astrocytoma ) Metastatic still images “Coronary , “PNET”, ixel amplit . These ima MRI scan ed on slice ti channels w The advan is Magnet e conducted ent patients w is written um xis 1, any l it and 𝑗 est the are . S ach b) e) of T1 and ude ges ner of ith ced om for ith on Image Anal Stereol 2018;37:249-275  257   Matlab15a software tool for execution of five algorithms. The number of solutions assigned initially is 100 and the maximum iteration number is assumed as 20. DISCUSSION In this paper, the performance of the proposed algorithm is explained in three sections: Evaluation of Algorithm performance, performance evaluation of proposed method in identifying tumour region and performance evaluation of proposed method for image compression. EVALUATION OF ALGORITHM PER- FORMANCE This section explains the performance of proposed method against other algorithms in terms of objective/maximum fitness function, mean value, standard deviation, and elapsed time or computational time. FITNESS FUNCTION: It explains how best a solution is fit to the problem. In this paper, Kapur’s entropy and Renyi’s entropy are considered as fitness function for effective and efficient image thresholding. Here the number of thresholds are 5 and are optimized with the proposed HBFOA-PSO. Table 1 shows the fitness value obtained with the proposed HBFOA-PSO is larger when compared with the other algorithms. MEAN AND STANDARD DEVIATION: All the optimization techniques are run more than 20 times and some randomness is involved in the execution process of the algorithm. Because of randomness, the algorithm never generates the same solution all times. The stability measuring parameter of the algorithm is mean and standard deviation. Mean value is the ratio of sum of maximum fitness value obtained with each run to total number of runs. Standard deviation is defined as a quantity expressing by how much the members of maximum fitness value differ from the mean value. From Table 1 it is clear that mean value and standard deviation of the proposed method is better than other algorithms. COMPUTATIONAL TIME It is measured in seconds and is total time taken by the algorithm to produce outcome or results. The proposed HBFOA-PSO algorithm computational time is little bit higher as compared with others because of cascading BFOA and PSO and is illustrated in Table 1. In comparison with the Renyi’s entropy and Kapur’s entropy, computational time of Renyi’s entropy is a little higher. PERFORMANCE EVALUATION OF PROPOSED METHOD IN IDENTIFYING TUMOUR REGION As disused in introduction section, in this section first it is explained how the proposed HBFOA-PSO is better in identification of brain tumour area as compared with other optimization techniques with maximizing Renyi’s entropy and Kapur’s entropy. This section explains visual clarity of tumour regions with the proposed method and other algorithms. For testing of HBFOA-PSO six images of different tumours in brain are chosen and all the images are in .jpg format. Fig. 4 to Fig. 9 show the optimal threshold images with optimized thresholds and which partition the image into clusters/groups and from these clusters one can clearly distinguish the tumour area and edema area in tumour affected brain image. Fig. 4 illustrates the threshold image of a female patient of age 42 and she is suffering with metastatic bronchogenic carcinoma brain tumour. From Fig. 4 it is observed that, visually tumour area and edema area are clearly partitioned with the proposed HBFOA-PSO compared with other algorithms. Input image shows that tumour is affected in left temporal region of the brain (high intensity region) but left hemisphere of the patient brain is totally inflamed as consequences of the all- encompassing surgery. For surgery radiologist may confuse in identifying the exact tumour location and it is possible with the HBFOA-PSO. Proposed HBFOA-PSO efficiency is tested on low grade Glioma brain tumour image and proved better compared to other algorithms and shown in Fig. 5. This Glioma brain tumour image is obtained from a patient of age 35 and tumour is identified in the left occipital area. Visual identification of tumour area and edema area are very clear with the proposed HBFOA-PSO. Fig. 6 shows the threshold image of patient of age 35 and was suffering with meningioma tumour. Reasonable occurrence of calcified pathology is reported in the right parietal convexity with a dural tail. The proposed HBFOA-PSO clearly identifies area of tumour in meningioma tumour image and this demonstrates the toughness and competence of the proposed HBFOA-PSO in identifying the tumour area of demanding clinical data. Fig.7 and Fig. 8 shows the threshold image of patient of age 32 and was suffering with high grade Astrocytoma tumour and these figures demonstrate the     efficiency a PSO. The H between righ rest area. I ventricular s of intense successfully In gener the patient when this le around six m is a difficult intensity lev images also as compared Primitiv massive tum Kapur’s Ent Input image Renyi’s Entr Fig. 4. Meta algorithms     nd effectiven BFOA-PSO t ganglio ca n spite of ystem, whic edema, the identified th al tumour b depends on vel is above onths. Ident task because el as shown i proposed HB to other algo e Neuro Ect our which ropy MFO opy static Bronc VIMA ess of the p visually show psular region the effacem h occurred d suggested e tumour reg ecomes canc the stage of four then lif ifying the tum there is no m n Fig. 9. Eve FOA-PSO s rithms. odermal Tum is mostly G hogenic Car LA KUMARI roposed HB s the tumou and thalamu ent of ipsil ue to the pre HBFOA-PSO ion. er and life ti tumour leve etime of pati our at early uch differe n with this k hows better r our (PNET appearing i WO cinoma affec G ET AL: Hyb 258 FOA- r area s and ateral sence has me of l and ent is stage nce in ind of esults ) is a n the ch pr ce Th co de un m ce pa po su H id ot PSO ted brain im rid Algorithm ildren at ag imitive or u lls try to ext ey appear s nsidered a s ad cells and equally. As assive aggre rebral hemis tient brain im st radio ther pratentorial t BFOA-PSO entifying PN her state-of-a B age and opti for Medical I e of below ndeveloped end to entire imilar to med ingle tumou around the t compared to ssive tumou pheres of th age of age 3 apeutic diagn umour. The e algorithm ET tumour a rt algorithms FOA mal threshol mage Compr 25 years an cells in the nervous syst uloblastoma r. PNET co umour fluid medullobla r which mo e brain. Fi and is obtai osis and wa ffectiveness is that it t early stage . HBFOA d images ob ession d it is due brain and th em of the br and were o ntain cysts s are distribu stoma, PNET stly affects g. 9 shows ned while do s suffering w of the propo works well as compared -PSO tained with f to ese ain. nce and ted is the the ing ith sed in to ive Image Anal S   Kapur’s Ent Input image Renyi’s Entr Fig. 5. Gliom Kapur’s Ent Input imag Renyi’s Entr Fig. 6. Meni   tereol 2018;3 ropy MFO opy a affected b ropy e MFO opy ngioma affec 7:249-275  G rain image an GW ted brain ima WO d optimal th O ge and optim 259 PSO reshold imag PSO al threshold B es obtained w BFO images obta FOA ith five algo A ined with five HBFOA rithms HBFOA-P algorithms -PSO SO     Kapur’s Ent Input image Renyi'sEntr Fig. 7. Astro Kapur’s Ent Input image Renyi’s Entr Fig. 8. Coro algorithms ropy MFO opy cytoma tumo ropy MFO opy nary T1 As VIMA G ur affected b G trocytoma tu LA KUMARI WO rain image a WO mour affecte G ET AL: Hyb 260 PSO nd optimal th PSO d brain ima rid Algorithm BFO reshold imag BF ge and optim for Medical I A esobtained w OA al threshold mage Compr HBFOA-P ith five algo HBFOA- images obt ession SO rithms PSO ained with five Image Anal S   Kapur’s Ent Input image Renyi’s Entr Fig. 9. Prim algorithms PERFO PROPO COMP This se PSO is effe PSNR, MSE PEAK S Peak sig nated in rec shows that e little low. T level MRI symbol MA MSE and in 20dB to 40d 𝑃𝑆𝑁𝑅 1 From th increased wi The bits image size pixels in com bpp and nu original ima number of t tereol 2018;3 ropy MFO opy itive Neuro RMANCE SED MET RESSION ction explain ctive in im , SSIM and W IGNAL TO nal to noise onstructed i ffect of noi he highest in brain image X. PSNR va general the r B and is calc 0𝑙𝑜𝑔 e above equa th the decrem per pixel (b (in terms of pressed ima mber of thr ge are repla hresholds Th 7:249-275  G Ectodermal t EVALUAT HOD FOR s how the p age compre PSNR. NOISE RAT ratio shows mage. A hig se on recons tensity level is 255 and lue is invers ange of PSN ulated using tion it is clea ent in MSE pp) is the ra bits) (Ĩ) an ge (ĨT). Belo esholds. All ced with opt =2, then 2 WO umour affec ION OF IMAGE roposed HB ssion in term IO the noise con h value of P tructed imag of the inpu is indicated ely proportio R value is be the Eq. (15). r that PSNR value. tio of compr d total numb w table show the pixels i imal thresho bits are enou 261 PSO ted brain im FOA- s of tami- SNR e is a t gray with nal to tween (15) value essed er of s the n the lds, if gh to re te 25 PS av H be is K th H al th m BFO age and opti present 2 th rms of bits) i 6*256). The Fig. 22 an NR and bi erage peak BFOA-PSO tter when co higher with apur’s entrop at reconstru BFOA-PSO gorithms and resholds obta arked on hist A mal threshol resholds. So s 256*256*2 refore bpp = d Fig. 23 sho ts per pixel signal to n is around 32 mpared to ot Renyi’s entro y. From Fig cted image is better Fig. 16 to ined with dif ogram of resp Number of thresholds (Th) 2 3 4 5 HBFOA-P d images ob compressed (since origin 256*256*Th ws the bar ch (bpp), from oise ratio o decibels and her algorithm py when co . 10 to Fig. 1 quality of as compar Fig. 21 sh ferent algori ective brain bpp 0.25 0.375 0.5 0.625 SO tained with f image size al image siz /256*256*8. art between these figu f the propo this value is s and its va mpared with 5 it is obser the propo ed with ot ow the optim thms and all tumour imag = Ĩ / ĨT ive (in e is the res sed far lue the ved sed her al are es.   VIMALA KUMARI G ET AL: Hybrid Algorithm for Medical Image Compression 262   MEAN SQUARE ERROR: It is the procedure of squaring the predictable quantities. It is the average error between the input and reconstructed image and the result is squared and is calculated using Eq. (16). 𝑀𝑆𝐸 ∑ ∑ 𝐼 𝐼 (16) The algorithm which gives lower value of mean squared error is the best algorithm. Lower value of MSE shows less difference between the input image and reconstructed image. In above equation M is size of the input image, I is original input image and Ĩ is reconstructed image or decompressed image. With the proposed method the value of MSE is lesser when compared to other algorithms. Fig. 24 and Fig. 25 show the rate distortion curve drawn between different images on x-axis and MSE on y-axis by considering bits per pixel (bpp) 0.625. These figures show the comparison of various MSE obtained with the various algorithms. MSE values are image compression measuring parameter which measures the deformation levels in the reconstructed image and this deformation levels are treated as error and is measured by taking pixel wise difference between input brain image and reconstructed image. As a whole proposed HBFOA-PSO has lower value of MSE when compared with others. STRUCTURAL SIMILARITY INDEX MEAS- URE (SSIM) PSNR and MSE values are measured with respect to intensity level of the input image and reconstructed image. Sometimes these two fail in measuring the reconstructed image visual quality. Sometimes PSNR value obtained with the technique may be high but visual quality is poor, so SSIM is introduced in this paper.The SSIM measures the similarity between input image and reconstructed image with separate luminance (L), contrast (C) and structure (S) components. SSIM of y and 𝑦 is calculated using following equation           IISIICIILyySSIM ~,~,~,,  (17) α, β and γ are the adjustable parameters which gives the relative importance of the three components and are equal to one in this paper for effortless calculation of SSIM.     21 2212 2 ~ 22 ~ 2 ~~ CC CC SSIM IIII IIII           (18) Where µI and µĨ are the mean value of the original image I and reconstructed image Ĩ, σI and σĨ are the standard deviation of original image I and reconstructed image Ĩ, σIĨ is the cross-correlation and C1 & C2 are constants which are equal to 0.065. The range of SSIM is -1 to +1 and SSIM value equal to one shows original image and reconstructed image is similar. The algorithm is said to be good if SSIM value is near around +1. Table 2 shows the SSIM of various methods with Renyi’s entropy and Kapur’s entropy and it shows proposed method SSIM is higher than other methods. 1 1 N I i i I N     (19)    Ii N i IiII IINncorrelatioCross ~1~ ~ 1 1      (20) WEIGHTED PSNR (WPSNR) The major advantage of PSNR is its simplicity in the calculation while major disadvantage is that it does not consider any of the human visual system (HVS) attributes. So there is need of WPSNR which incorporate HVS parameters. WPSNR is HVS-based method and more accurate than PSNR. The WPSNR uses the principle of redundancy of the human eye toward high frequency components in images. The human perception of vision is less sensitive to edges than smooth areas. The WPSNR is nothing but PSNR weighted by the HVS parameter (Navas et al., 2011). The WPSNR in dB is expressed as WPSNR 10log (21) Where NVF is noise visibility function and defined as NVF norm (22) Where, δblock is the standard deviation of pixels having a specific size (8×8). In smooth regions, the value of NVF is near to zero and in the regions with edges and texture it is near to unity. From Table 2 it is observed that proposed method is better in WPSNR compared to other methods. Image Anal S   Kapur’s Ent Input image Renyi’s Entr Fig. 10. Astr Kapur’s Ent Input image Renyi’s Entr Fig. 11. Co tereol 2018;3 ropy MFO opy ocytoma tum ropy MFO opy ronary T1 As 7:249-275  G our affected GW trocytoma tum WO brain image O our affected 263 PSO and decompr PSO brain image a BF essed image BFO nd decompre OA obtained wit A ssed images o HBFOA-P h five algorit HBFOA-PSO btained with SO hms five algorithms     Kapur’s Ent Input image Renyi’s Entr Fig. 12. Glio Kapur’s Ent Input image Renyi’s Entr Fig. 13. Met ropy MFO opy ma affected ropy MFO opy astatic Bronch VIMA GW brain image G ogenic Carci LA KUMARI O and decompr WO noma affected G ET AL: Hyb 264 PSO essed images PSO brain image rid Algorithm BFO obtained wi BF and decompr for Medical I A th five algori OA essed images mage Compr HBFOA-PS thms HBFOA-PS obtained with ession O O five algorithms Image Anal S   Kapur’s Ent Input image Renyi’s Entr Fig. 14. Pri algorithms Kapur’s Ent Input image Renyi’s Entr Fig. 15. Men tereol 2018;3 ropy MFO opy mitive Neuro ropy MFO opy ingioma affe 7:249-275  G Ectoderma GW cted brain im WO l tumour aff O age and dec 265 PSO ected brain PSO ompressed im BF image and d BFO ages obtaine OA ecompressed A d with five a HBFOA-P images obt HBFOA-P lgorithms SO ained with f SO ive   VIMALA KUMARI G ET AL: Hybrid Algorithm for Medical Image Compression 266   Kapur’s Entropy               MFO GWO PSO BFOA HBFOA-PSO Renyi’s Entropy Fig. 16. Optimal thresholds on Histogram of Astrocytoma tumour affected brain image with five algorithms Kapur’s Entropy MFO GWO PSO BFOA HBFOA-PSO Renyi’s Entropy Fig. 17. Optimal thresholds on Histogram of Coronary T1 Astrocytoma tumour affected brain image with five algorithms 0 50 100 150 200 250 300 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 x 104 hi st og ra m ,T hr es ho ld s histogrm Thresholds 0 50 100 150 200 250 300 0 0.2 0.4 0.6 0.8 1 .2 .4 .6 .8 2 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 2 4 6 8 1 2 4 6 8 2 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 x 104 g histogrm Thresholds 0 50 100 150 200 250 300 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 2 4 6 8 1 2 4 6 8 2 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 x 104 g , histogrm Thresholds 0 50 100 150 200 250 300 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 x 104 g , histogrm Thresholds 0 50 100 150 200 250 300 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 x 104 g , histogrm Thresholds 0 50 100 150 200 250 300 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 x 104 g , histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 i l i t it histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 .5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 g histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds H ist og ra m , T hr es ho ld s  Pixel Intensity H ist og ra m , T hr es ho ld s  Pixel Intensity H ist og ra m , T hr es ho ld s  H ist og ra m , T hr es ho ld s  Pixel Intensity Pixel Intensity Image Anal Stereol 2018;37:249-275  267   Kapur’s Entropy                 MFO GWO PSO BFOA HBFOA-PSO Renyi’s Entropy Fig. 18. Optimal thresholds on Histogram of Glioma affected brain image with five algorithms Kapur’s Entropy MFO GWO PSO BFOA HBFOA-PSO Renyi’s Entropy Fig. 19. Optimal thresholds on Histogram of Metastatic Bronchogenic Carcinoma affected brain image with five algorithms 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 g histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 0000 2000 4000 6000 histogrm Thresholds 0 50 100 150 200 250 300 0 000 000 000 000 000 000 000 000 histogrm Thresholds 0 50 100 150 200 250 300 0 000 000 000 000 000 000 000 000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 histogrm Thresholds 0 50 100 150 200 250 300 0 000 000 000 000 000 000 000 000 000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 0000 2000 4000 6000 8000 histogrm Thresholds 0 50 100 150 200 250 300 0 000 000 000 000 000 000 000 000 000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 0000 2000 4000 6000 8000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 0000 2000 4000 6000 8000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 0000 2000 4000 6000 8000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 0000 2000 4000 6000 8000 histogrm Thresholds 0 50 100 150 200 250 300 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 histogrm Thresholds H ist og ra m , T hr es ho ld s  H ist og ra m , T hr es ho ld s  H ist og ra m , T hr es ho ld s  H ist og ra m , T hr es ho ld s  Pixel Intensity Pixel Intensity Pixel Intensity Pixel Intensity   VIMALA KUMARI G ET AL: Hybrid Algorithm for Medical Image Compression 268   Kapur’s Entropy MFO GWO PSO BFOA HBFOA-PSO Renyi’s Entropy Fig. 20. Optimal thresholds on Histogram of Primitive Neuro Ectodermal tumour affected brain image with five algorithms Kapur’s Entropy MFO GWO PSO BFOA HBFOA-PSO Renyi’s Entropy Fig. 21. Optimal thresholds on Histogram of Meningioma affected brain image with five algorithms     0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 10 4 g histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 10 4 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 10 4 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 .5 1 .5 2 .5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 .5 1 .5 2 .5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 10 4 g histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 10 4 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 10 4 histogrm Thresholds 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 x 104 g , histogrm Thresholds H ist og ra m , T hr es ho ld s  H ist og ra m , T hr es ho ld s  H ist og ra m , T hr es ho ld s  H ist og ra m , T hr es ho ld s  Pixel Intensity Pixel Intensity Pixel Intensity Pixel Intensity Image Anal Stereol 2018;37:249-275  269   Table 1. Evaluation of Fitness, Mean, Standard deviation and Elapsed time of five methods for brain images Image    Optimization Technique Fitness value Mean Standard deviation Elapsed time (sec) Kapur Renyi Kapur Renyi Kapur Renyi Kapur Renyi Meningioma GWO 9.8524 16.172 10.465 16.904 0.08711 0.117654 28.931 10.809 MFO 10.391 16.631 10.483 16.922 0.0001 3.61E-15 9.0867 7.2405 PSO 10.483 16.913 10.055 15.99 0.19127 0.410305 19.165 9.2245 BFOA 10.485 16.922 10.485 16.913 5.42E-15 1.45E-14 8.0601 7.0205 HBFOA-PSO 10.537 16.968 9.8524 16.172 3.61E-15 3.61E-15 39.543 57.999 PNET GWO 10.81 15.971 10.81 15.971 0.03428 0.144739 20.027 16.979 MFO 11.149 16.912 11.088 16.122 5.42E-15 3.61E-15 10.23 7.1725 PSO 11.176 17.015 11.149 17.015 0.0982 0.569382 55.656 9.8013 BFOA 11.183 17.075 11.172 17.02 5.42E-15 1.08E-14 13.103 7.6018 HBFOA-PSO 11.187 17.086 11.176 17.075 5.42E-15 7.23E-15 40.608 59.407 Metastatic GWO 12.191 19.011 12.191 19.011 0.03701 0.038206 18.119 15.381 MFO 12.518 19.704 12.518 19.419 9.03E-15 1.08E-14 21.278 10.68 PSO 12.546 19.711 12.34 19.711 0.12774 0.22731 19.185 94.153 BFOA 12.55 19.762 12.535 19.762 9.03E-15 0.0909 9.7452 9.4357 HBFOA-PSO 12.551 19.765 12.551 19.749 1.81E-15 0.09090 38.520 48.261 Glioma GWO 12.002 19.041 12.002 19.041 0.03799 0.05544 17.759 12.135 MFO 12.36 19.634 12.159 19.634 5.42E-15 1.08E-14 12.25 7.8565 PSO 12.363 19.658 12.368 19.658 0.12408 0.20444 38.2 9.9885 BFOA 12.394 19.824 12.36 19.501 9.03E-15 3.61E-15 9.8131 17.692 HBFOA-PSO 12.395 19.835 12.363 19.805 0.0909 3.61E-15 45.879 52.848 Coronary T1 Astrocytoma GWO 10.881 16.965 10.881 16.965 0.03465 0.063441 25.367 16.936 MFO 11.027 18.248 10.953 18.248 0.0075 3.61E-15 8.4109 9.0397 PSO 11.102 18.268 11.027 18.268 0.12238 0.341585 17.462 10.561 BFOA 11.156 18.273 11.102 17.681 3.61E-15 7.23E-15 12.135 8.3687 HBFOA-PSO 11.165 18.348 11.136 18.3 1.81E-15 3.61E-15 36.64 58.186 Astrocytoma GWO 10.736 17.244 11.011 18.052 0.02483 0.078797 17.74 21.74 MFO 11.022 17.635 11.024 17.825 1.81E-15 3.61E-15 7.577 8.0546 PSO 11.024 17.825 10.861 17.543 0.13426 0.424639 16.351 13.634 BFOA 11.024 18.04 11.024 17.635 3.61E-15 3.61E-15 14.347 7.9109 HBFOA-PSO 11.025 18.091 10.736 17.244 1.81E-15 7.23E-15 50.755 65.194         VIMALA KUMARI G ET AL: Hybrid Algorithm for Medical Image Compression 270   Table 2. Evaluation of SSIM and WPSNR of five methods for brain images Image    Optimization Technique SSIM WPSNR (dB) Kapur Renyi Kapur Renyi Meningioma GWO 0.658670 0.886195 32.04823 15.609031 MFO 0.667427 0.904109 32.058787 15.937439 PSO 0.667626 0.904584 32.063169 16.052494 BFOA 0.667959 0.904651 32.069061 18.125513 HBFOA-PSO 0.667986 0.910922 32.108383 19.708357 PNET GWO 0.6827744 0.832106 19.053239 18.153405 MFO 0.7056383 0.8752121 23.705779 18.531831 PSO 0.7226091 0.8762945 24.686436 18.943262 BFOA 0.7412038 0.8776707 27.591976 19.081611 HBFOA-PSO 0.7586052 0.8814224 27.983886 20.746489 Metastatic GWO 0.6717987 0.7295003 27.416015 17.666669 MFO 0.6747454 0.7368438 27.788892 26.471868 PSO 0.6803707 0.7369216 27.863482 26.496026 BFOA 0.6814665 0.740371 28.048519 26.500085 HBFOA-PSO 0.6953937 0.8511379 28.21856 26.725894 Glioma GWO 0.5983156 0.6936872 22.066536 16.974562 MFO 0.6161177 0.6949095 27.936855 17.619167 PSO 0.6312582 0.8396407 28.838197 18.795318 BFOA 0.6353784 0.841043 29.477103 28.513004 HBFOA-PSO 0.7363141 0.8439056 30.557768 28.723952 Coronary T1 Astrocytoma GWO 0.6350157 0.8571572 22.77016 17.628194 MFO 0.6386655 0.8581259 29.271111 17.852049 PSO 0.6450002 0.8588663 29.412006 18.156052 BFOA 0.6486089 0.8625398 29.809665 19.952415 HBFOA-PSO 0.8092489 0.8648337 30.562705 20.525345 Astrocytoma GWO 0.6304807 0.7852364 22.866723 17.079014 MFO 0.6307575 0.7952396 24.701479 17.632169 PSO 0.6566649 0.8541937 27.212897 18.608922 BFOA 0.7165008 0.8649524 29.294284 19.586243 HBFOA-PSO 0.7276255 0.8665126 29.294394 19.822887 Image Anal Stereol 2018;37:249-275  271   Fig. 22. Peak Signal-to-Noise Ratio of brain images obtained with five algorithms using Renyi Entropy 0 5 10 15 20 25 30 35 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) PNET  GWO MFO PSO BFOA HBFOA‐PSO 0 5 10 15 20 25 30 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Meningioma GWO MFO PSO BFOA HBFOA‐PSO 0 5 10 15 20 25 30 35 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Metastatic GWO MFO PSO BFOA HBFOA‐PSO 0 5 10 15 20 25 30 35 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Glioma  GWO MFO PSO BFOA HBFOA‐PSO 0 5 10 15 20 25 30 35 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Coronary T1 Astrocytoma  GWO MFO PSO BFOA HBFOA‐PSO 0 5 10 15 20 25 30 35 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Astrocytoma  GWO MFO PSO BFOA HBFOA‐PSO   VIMALA KUMARI G ET AL: Hybrid Algorithm for Medical Image Compression 272   Fig. 23. Peak Signal-to-Noise Ratio of MRI brain images obtained with five algorithms using Kapur’s Entropy 0 10 20 30 40 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) PNET GWO MFO PSO BFOA HBFOA‐PSO 0 10 20 30 40 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Coronary T1 Astrocytoma GWO MFO PSO BFOA HBFOA‐PSO 0 10 20 30 40 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Astrocytoma   GWO MFO PSO BFOA HBFOA‐PSO 0 10 20 30 40 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Metastatic  GWO MFO PSO BFOA HBFOA‐PSO 0 10 20 30 40 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Glioma  GWO MFO PSO BFOA HBFOA‐PSO 0 5 10 15 20 25 30 bpp=0.25 bpp=0.375 bpp=0.5 bpp=0.625 PS NR  (d B) Meningioma  GWO MFO PSO BFOA HBFOA‐PSO Image Anal Stereol 2018;37:249-275  273   Fig. 24. Rate distortion curve for different images with Kapur’s Entropy at bits per pixel (bpp) 0.625 Fig. 25. Rate distortion curve for different images with Renyi Entropy at bits per pixel (bpp) 0.625 CONCLUSIONS Numerous experiments and design benchmarks, based on the comparison of HBFOA-PSO with other algorithms, established that it is highly efficient with an almost exponential convergence rate for solving optimization problems. Therefore, HBFOA-PSO algorithm is a trustworthy technique for solving complicated problems in the process of image compression. This study enumerates how the HBFOA- PSO uses the Renyi’s entropy and Kapur’s entopy for optimal thresholds, hence better identification of tumour regions and better reconstructed visual image quality. The HBFOA-PSO algorithm is implemented for compression of MRI brain images. All the results of five experiments established that the HBFOA-PSO algorithm will increase the fidelity of reconstructed images compared to other four methods. Moreover the results of experiments established that the proposed HBFOA-PSO algorithm is capable of identifying the brain tumour regions and have better PSNR, MSE, SSIM and WPSNR when compared with other algorithms. Finally, the performance of the proposed HBFOA-PSO algorithm is found to be better with Renyi’s entropy when compared with the Kapur’s entropy.   0 20 40 60 80 100 120 M SE GWO MFO PSO BFOA HBFOA‐PSO 0 10 20 30 40 50 60 70 M SE GWO MFO PSO BFOA HBFOA‐PSO   VIMALA KUMARI G ET AL: Hybrid Algorithm for Medical Image Compression 274   REFERENCES Anitha V, Pallikonda R, Vishnuvarthanan G, Yudong Z, Arunprasath T (2017). An automated hybrid approach using clustering and nature inspired optimization technique for improved tumour and tissue segmenta- tion in magnetic resonance brain images, Appl Soft Comput, 57: 399–26. Badri NS, Veerakumar T, Esakkirajan SA (2016). Tumour or abnormality identification from magnetic resonance images using statistical region fusion based segmenta- tion, Magn Reson Imaging, 34(9): 1292-304. Chen-Kuei Y, Wen-Hsiang. T (2015). Color image compression using quantization, thresholding, and edge detection techniques all based on the moment- preserving principle, Pattern Recog Lett, 19: 205-15. Diego O, Salvador H, Erik C, Gonzalo P, Omar A, Jorge G (2017). Cross entropy based thresholding for magnetic resonance brain images using Crow Search Algorithm, Expert Syst Appl, 79: 164–80 GS Rao, PMK Prasad, DYV Prasad (2016). Performance Analysis of orthogonal and Biorthogonal wavelets for Edge detection of X-ray Images, Procedia Comput Sci, 87:116-21. Irem EK, Ayca Çi, Emine GS, Turgay I (2017). PCA based clustering for brain tumour segmentation of T1W MRI images, Comput Meth Prog Bio, 140: 19- 28. Ivan C, Iker G (2017). MRI segmentation fusion for brain tumour detection, Inform Fusion, 36: 1–9. Kaur LS Gupta RC, Chauhan SCS (2007) Medical ultrasound image compression using joint optimization of thresholding quantization and best-basis selection of wavelet packets, Digit Signal Process, 17: 189-98. Kaveh Ahmadi, Ahmad YJ, Ezzatollah S (2015). An efficient compression scheme based on adaptive thresholding in wavelet domain using particle swarm optimization, Signal Process-Image, 32: 33–39. Kennedy J, Eberhart RC (1995). A new Optimizer using Particle Swarm Theory, in Proceedings of Sixth International Symposium on Micro Machine and Human Science, Nagoya, Japan, 6: 39-43. Lubna F, Adeel Y and Rana HR (2017). Adaptive stochastic segmentation via energy-convergence for brain tumour in MR image., J Vis Commun Image R, 46: 303–11. Navas KA, Gayathri Devi KG, Athulya MS, Anjali V (2011). MWPSNR: A new image fidelity metric, IEEE Recent Advances in Intelligent Computational Systems (RAICS), 627-32. Nooshin N, Miroslav K (2017). Automatic tumour segmentation in single-spectral MRI using a texture- based and contour-based algorithm, Expert Syst Appl, 77: 1–10. Passino KM (2002). Biomimicry of bacterial foraging for distributed optimization and control, IEEE Control Systems Magazine, 22: 52–67. Prasanna S, Carrye W, Jerry Y (1997). Threshold Selection Using Renyi's Entropy, Pattern Recogn, 30(1): 71-84. Prashant S, Ioana M (2003). Selective Thresholding in Wavelet Image Compression, Wavelets and Signal Processing Part of the series Applied and Numerical Harmonic Analysis, 2: 377-81. Rajeswari R (2012). Type-2 Fuzzy Thresholded Bandlet Transform for Image Compression, Procedia Engineer, 38: 385-90. Rajinikanth V, Suresh CS, Steven LF, Nachiappan S (2017). Entropy based segmentation of tumour from brain MR images – a study with teaching learning based optimization, Pattern Recog Lett, 94: 87–95. Ramakrishnan T, Sankaragomathi B (2017). A professional estimate on the computed tomography brain tumour images using SVM-SMO for classifica- tion and MRG-GWO for segmentation, ttern Recog Lett, 15:163–71. Salleh. MFM, Soraghan J (2007). A New Multistage Lattice Vector Quantization with Adaptive Subband Thresholding for Image Compression, EURASIP Journal on Advances in Signal Processing December, 2: 23-34. Sathya PD, Kayalvizhi R (2011). Optimal multilevel thresholding using bacterial foraging algorithm, Expert Syst Appl, 38: 15549–64. Sathya PD, Kayalvizhi. R (2011). Amended bacterial foraging algorithm for multilevel thresholding of magnetic resonance brain images, Measurement, 44: 1828–48. Siraj. S (2015). Comparative study of Birge–Massart strategy and unimodalthresholding for image compression using wavelet transform, Optik, 126: 5952-55. Solmaz A, Farshad T (2017). Detection of brain tumour in 3D MRI images using local binary patterns and histogram orientation gradient. Neurocomputing, 219: 526–35. Sudip Kumar A, Jamuna KS, Dipak K, Basu MN (2015). Conditional spatial fuzzy C-means clustering algorithm for segmentation of MRI images, Appl Soft Comput, 34: 758-69. Tahere IM, Mohammad RKM (2009). ECG Compression with Thresholding of 2-D Wavelet Transform Coefficients and Run Length Coding, Eur J Sci Res, 27: 248-57. Tony FC, Zhou H (2007). Total Variation Wavelet Thresholding, J Sci Comput, 32(2): 315-41. Image Anal Stereol 2018;37:249-275  275   Yudong Z, Lenan W (2011). Optimal multi-level thresholding based on maximum Tsallis entropy via an artificial bee colony approach, Entropy, 13(4): 841–59. Yuhong L, Fucang J, Jing Q (2016). Brain tumour segmentation from multimodal magnetic resonance images via sparse representation, Artif Intell Med, 73: 1-13.