Home Computer Science



Adaptive Regularized Gaussian Kernel FCM for the Segmentation of Medical Images: An Artificial IntelligenceBased IoT Implementation for Teleradiology NetworkINTRODUCTIONSegmentation method is exploited for the extraction of desired region of interest (ROI) and in medical image processing, its role is pivotal in the delineation of anatomical organs and anomalies like tumor and cyst. There is no universal algorithm for various modalities, and the selection of segmentation technique relies on the type of imaging modality and ROI. Grouping of segmentation techniques relies on the nature of evolution and, in general, it is classified into semiautomatic and fully automatic. The semiautomatic algorithm requires human intervention: the discrete positioning of points in the level set model [1], selection of foreground, and background seed region in graph cut [2]. The big data analytics and Internet of Things (IoT) trends influence healthcare in the radiology sectors for the classification in efficient diagnosis [3]. A realtime mobile camera terminal captures the skin images that interact with the remote datacenter with a deep learning model, which improvises the learning model and predicts the skin disease classification and psychological depression [4]. For the analysis of medical imaging, the Raspberry Pi with sensors is used to collect data from the clinical environment by automatically logging telephone calls [5]. An ARM9based processor was used for the intelligent processing of medical images, and the processed image was securely transferred with LiFienabled IoT [6]. The authors have reviewed the IoT applications in the Ehealthcare by different types of sensors: medical big data management [7]. The author proposed a polynomial time algorithm for the efficient cloud data transfer to medical IoT devices with energyefficient dynamic packet downloading of the medical images. This method adaptively changes power/energy at each access point with buffer stability [8]. The deep learning framework was proposed with Caffe or Tensor flow for training the machine learning model with medical images in a big data method. The trained model is used in mobile application for analysis and diagnosis of disease through cloud for Internet of Medical Imaging Thinks (IoMIT) [9]. The author introduced an edge computing device between the IoT device and cloud network for the effective analysis of CT brain images, especially for strokes patient. This model uses Adaptive Boosted Decision Trees (ABDT) and intelligent classifiers for the classification of cerebrospinal fluid (CSF), white matter (WM), gray matter (GM), ischemic stroke, hemorrhagic stroke, calcification, and bone for the better prediction of stroke recovery [10]. Fuzzy СMeans (FCM) is a soft clustering algorithm that groups the pixels with respect to fuzzy membership function. The random initialization of centroids, the prior setting of cluster number, and stuck at local minima are the discrepancies of classical FCM algorithm [11]. In general, the CT and MR images are corrupted by Gaussian and Rician noise, respectively. Moreover, appropriate filtering is required in the realtime scenario before any other image processing operations like segmentation, classification, and compression to improve accuracy [12]. Fuzzy clustering finds its application in medical and remote sensing areas [13, 14]. The traditional FCM initializes the cluster centroids arbitrarily, and it stucks in local optima regularly. Initially, Dunn implemented FCM and it is modified by Hathaway and Bezdek [15, 16], whose objective function is given in equation (1.1).
where m is the fuzzy weighted exponential factor (m > 1), d_{a}p is the Euclidean distance, and U_{a}p indicates the membership function. The Euclidean distance is represented in equation (1.2).
The membership function satisfies the following constraint and is represented in equation (1.3):
Conventional FCM is subtle to noise due to the lack of local pixel information in objective function. The spatial FCM takes into account the neighborhood pixel information [17], whose objective function is expressed in equation (1.4):
у controls the spatial information of the neighboring pixels with 0 > f > . N_{a} is the group of pixels and N_{R} is the cardinality of N_{a}. The computational complexity of spatial fuzzy cmeans (SFCM) is high and hence improved SFCM was proposed, which replaces ^l,_{r(Na}dp with d^{2}^mean(^x_{a},Vp^ [18] and is represented in equation (1.5):
More time domain information is incorporated in objective function by replacing mean [x_{a}) by vector median of neighbours around x_{a} [19] and is expressed in equation (1.6):
where V_{m}(x_{a}) is the vector median of image elements in the predefined kernel (usually 3x3). The SFCM and enhanced spatial fuzzy cmeans (ESFCM) require the estimation of parameters, and this problem was solved by fuzzy local information cmeans (FLICM) algorithm [20]. The objective function of FLICM is represented in equation (1.7):
G_{a}p is the fuzzy factor that incorporates both time domain and grayscale information of neighboring image elements. In every iteration, determination of fuzzy factor is needed, and hence FLICM algorithm is slow and also there is a loss of fine details of image due to the smoothing factor. The kernel weighted fuzzy local information cmeans (KWFLICM) technique incorporates a weighted fuzzy factor in the objective function [21] and is represented in equation (1.8):
where
The computation cost increases due to the tradeoff weighted fuzzy factor and there is a loss of fine details of the image. Though FLICM and KWFLICM improve the segmentation result by the inclusion of time domain and grayscale information, the objective function is not minimized to an optimum value. Hwang modified the FCM by replacing the conventional fuzzy membership function [22]. The expression U_{a}p was replaced by a_{a}p and is expressed in equation (1.9) as follows:
The type 2 FCM does not produce satisfactory output in the case of complex ROI patterns in the image. The objective function of type 2 FCM is represented in equation (1.10):
It will terminate at a point, at which the previous and updated membership satisfy the following criteria represented in equation (1.11):
The novel variant of the FCM algorithm is institutionistic fuzzy cmeans (IFCM) based on institutionistic fuzzy set [23]. Section 1.2 highlights the proposed methodology, Section 1.3 describes the results and discussion, and Section 1.4 constitutes the conclusion. 
<<  CONTENTS  >> 

Related topics 