Multi-objective Metaheuristics with Intelligent Deep Learning Model for Pancreatic Tumor Diagnosis

Initially, the Weiner filter (WF) is employed as a pre-processing tool for pancreatic tumor diagnosis. The WF is a new linear filter proposed by Norbert Wiener in 1942. The WF technique is chosen over other methods due to following reasons. It executes an optimum trade-off among inverse filtering and noise smoothening. It will remove the additive noise and overturns the blurring concurrently.

The aim is to minimize the MSE amongst the foreseeable actual output and noise free signals. Hence, it cannot be adopted and executed continuously in frequency domain. Generally, consider f (x, y) represents the input image and g (x, y) denotes the degraded image with few points spread function h (x, y) and additive noise (x,  y). Thus, in spatial domain, the blurred image can be expressed by: g ( x , y ) = H ( x , y ) * f ( x , y ) + η ( x , y ) ,(1) whereas * means two dimension convolutional, H (x, y) be the blurring function and additive noise η (x, y) often indicates uniform noise, Gauss white noise, etc. The WF provides noise and image as random model, and the aim is to find an assessment f ˆ of original images f (x, y) that MSE is minimum. The optimization problem can be expressed as follows: min e 2 = E { ( f - f ˆ ) 2 } ,(2)

Let E be the arithmetical form. In the frequency domain, the optimized solutions are given by: F ˆ ( u , v ) = H * ( u , v ) ( | H ( u , v ) | 2 + S η ( u , v ) / S f ( u , v ) ) ,(3)

In which H^* (u,  v) means the difficult conjugate of H (u,  v), S _η  (u,  v) determines the power spectrum of noise and S _f  (u, v) signifies the power spectrum of original image. While (S _η  (u,  v)/S _f  (u,  v)) is higher, the WF is lower, hence the frequency can be neglected.

Kapur’s image segmentation is first developed in 1985 for segmenting the preprocessed images through the entropy of histogram [20]. A familiar entropy-based thresholding method is the Kapurthresholding method. The criterion to elect an appropriate threshold is maximizing Kapur’s entropies depending upon gray-level histogram. Such approaches find the optimum T that maximizes the entire entropy. Let represents a vector of the image threshold. Subsequently, Kapur’s entropy would be represented as follows. J max = f _ kapur ( T ) = ∑ j = 1 k H j C for C { 1 , 2 , 3 }(4)

Commonly, every entropy is calculated individually according to the certain t value. For multilevel thresholding problems, it could be given by; H 1 C = ∑ j = 1 t 1 Ph j C ω 0 C ln ( Ph j C ω 0 C ) , H 2 C = ∑ j = t l + 1 t 2 Ph j C ω 1 C ln ( Ph j C ω 1 C ) , ⋮ H k C = ∑ j = t k + 1 L Ph j C ω k - 1 C ln ( Ph j C ω K - 1 C ) ,(5) whereas Ph j C represents the likelihood distribution of the intensity level and ω 0 C , ω 1 C , … , ω k - 1 C likelihood existence for k level. In order to optimally choose the threshold value, the BMO algorithm is applied.

The barnacles are microorganisms which assign themselves to objects from the water. The long penis is an important feature. Its mating group contains every neighbor as well as competitor in the influence of its penis. The barnacle mating optimizing was simulated as the mating procedure of barnacles. With simulating 3 procedures (for instance, initialization, selection model, and reproduction), the practical optimized issue was resolved. A briefly are explained as follows [21]:

Initially, it can be supposed that the candidate solution has been barnacle, in which the matrix of population is written utilizing in Equation (6). The estimation of population and sorting model is completed for locating the optimum solution until at the top of X. Afterward, the parent that exists mated were elected in Equations (7) and (8). X = [ x 1 1 ⋯ x 1 n ⋮ ⋱ ⋮ x N 1 ⋯ x N n ](6) barnacle _ d = randperm ( N )(7) barnacle _ m = randperm ( N )(8) where N implies the amount of barnacle population, n refers the amount of control variables, and barnacle _ d and barnacle _ m demonstrate the parents to be mated.

As there are no detailed formulas for deriving the reproduction procedure of barnacles, the BMO emphasizes the genotype frequency of parents for producing the offspring dependent upon Hardy–Weinberg rule. It can be worth importance that the length of its penises (pl) roles a vital play in determining the exploitation as well as exploration models. Considering pl = 7, it could be realized that barnacle #1 can only mate with most barnacles #2-#7. When the selective of barnacles that exist mated in the range of pl of Dad barnacle, the exploitation model has happened. Equation (9) is presented for producing novel variables of offspring in the barnacle parent. x i N _ new = px barnacle _ d N + qx barnacle _ m N(9) where p refers the normally distributed arbitrary number amongst 0 and 1, q = (1 - p), x barnacle _ d N and x barnacle _ m N signify the variables of Dad and Mum barnacles which are elected from Equations (7) and (8). p and q imply the genotype frequency of Dad and Mum barnacles from the novel offspring. Figure 2 showcases the flowchart of BMO technique.

When barnacle #1 elects barnacle #8, it can be over restrict. Therefore, the normal mating procedure doesn’t happen. Here, the offspring was formed by sperm cast model. In BMO, the sperm cast has been assumed as the exploration method that is written as:

OPEN IN VIEWER

Fig. 2

Flowchart of BMO.

x i n _ new = rand ( X ) × x barnacle _ m n(10)

Where, rand (X) refers the arbitrary number amongst 0 and 1.

It could be realized in Equation (10) which the novel offspring was created by Mum barnacle as it achieves the sperms which are released as to water by another barnacle away. In the iteration, the place of barnacle was upgraded based on Equation (9) or Equation (10). Lastly, the BMO is determined for approximating the global optimal to optimizedissues.

A dense CNN, also called as DenseNet is used for extraction of the features of the segmented image. All layers affect each other layers in a feed forward way. A DenseNet block includes 5 layers. The initial 4 layer denotes dense layers, and the latter is a transition layer. When the growth rate value is (k) for all the layers, it is four for these dense blocks. In the transition layer, it consists of 2×2 average pool using a 1×1 conv and a stride of two. In the dense layer, it consists of 3×3 and 1×1 conv using astride of 1.

The first layer of DL network captures generic features, whereas last one focuses on task specific features. Following weights from the ImageNet to pre-trained the model, they leverage prior learning of generic feature capture in the first layer and reduces the training data requirement. The architecture employed in these studies employ model parameter better and utilize lesser data to train compared to other frameworks and have a low tendency for overfitting to training the data [22]. DenseNet-169 model is depending upon the DenseNet structure. While the classical convolution network using L layers has L connection - one among every layer and its succeeding layers— DenseNet has L (L + 1)/2 direct connection. For all the layers, the feature map of each preceding layer is applied as an input, make DenseNets highly compact model with higher features reutilize over the network.

In the training procedure, image X from the training datasets is fed to the M model, i.e., determined using its parameter β. The output of the models is related to the training labels which correspond to the input images, the losses are calculated, and the models are upgraded by a novel parameter. In order to initiate the training procedure, models are started with weight pretrained on ImageNet, a huge CV database.

Figure 3 depicts the framework of DenseNet-169 model. In the DenseNet-169 architecture, the last fully connected layer was eliminated, in its defect, a 256 node fully connected layer was created, then a 128 node FLC and at the end a 10 class FLC with softmax activation for the output. All but the final layer used ReLU activation.

OPEN IN VIEWER

Fig. 3

Structure of DenseNet-169 model.

Afterward pretraining, the weights are fine-tuned by the training data. Pretraining and fine-tuning, decrease the time taken for the models to converge. Also, it assists in normalizing the method for reducing overfitting. In the iterative training procedure, a softmax activation function was employed on the estimations beforehand the loss was computed. Cross entropy has been applied as the loss function to be optimized.

The FSVM model receives the feature vectors as input and performs classification process. In standard SVM, every data point is assumed with equivalent significance and allocated the similar penal parameter from their objective functions. But, in several real-world classifier applications, any instance points like the outlier/noise, could not be exactly allocated to one of these 2 classes, and all instance points don’t have a similar meaning to decision surface [23]. The fuzzy membership for all instance points is established such that distinct instance points can generate various influences to the structure of decision surface.

Let the training instances are S = { ( x i , y i , s i ) , i = 1 , … , N } ,(11) where x_i ∈ Rⁿ implies the n dimensional instance point, y_i∈ { - 1, + 1 } refers their class label, and s _i  (i = 1, …, N) is a fuzzy membership that fulfil sσ ⩽ s_i ⩽ 1 with appropriately small constant σ > 0. The quadratic optimize issue to classifier is assumed as: min 1 2 w T w + C ∑ i 1 l s i ξ i(12) s . t . y i ( w T x i + b ) ⩾ 1 - ξ i , ξ i ⩾ 0 , i = 1 , … , l , where w refers the normal vector of separate hyperplane, b refers the bias term, and C implies the parameter that is to be defined previously for controlling the trade-off amongst the classification margin as well as the cost of misclassification error. While s_i is the attitude of equivalent point x_i nearby one class and slack variables ξ_i are measure of error, afterward the term s _i ξ _i is assumed that the measure of error with various weights. It can be noticeable the superior s_i is the most significantly the equivalent point has been preserved; the lesser than s_i is the less significantly the equivalent point has been preserved; therefore, distinct input points are generatedwhich helps in the learning of decision surface. So, the FSVM is defined as a further robust hyperplane by maximized the margin by letting any misclassification of lesser vital points. Figure 4 illustrates the hyperplane of SVM model.

OPEN IN VIEWER

Fig. 4

Hyperplane of SVM.

For resolving the FSM optimum issue, (12) has been altered as to the following dual issue by presenting Lagrangian multipliers α_i:

max ∑ i = 1 N α i - 1 2 ∑ i = 1 N ∑ j = 1 N α i α j y i y j x i x j(13) s . t . ∑ i = 1 N y i α i = 0 , 0 ⩽ α i ⩽ s i C , i = 1 , … , N .

Related to the typical SVM, the beyond statement only has little variance that is upper bound of the values of α _i . With resolving this dual issue from (13) for optimum α_i, w and b are improved in a similar manner as in the classic SVM. In order to tune the parameters of the FSVM model, the BMO algorithm is applied and thereby raises the classification performance.