Images super-resolution by optimal deep AlexNet architecture for medical application: A novel DOCALN 1

Abstract

In the last decade, numerous researches have been focused on Image Super-Resolution (SR); this recreation or improvement model is vital in different research areas. Recently, deep learning algorithm finds useful to advance in the resolution of the medical output. Here, we devise a novel Deep Convolutional Network model along with the optimal learning rate of the Rectified Linear Unit (ReLU) intended for Medical Image Super-Resolution (MISR). For getting the optimal values of Deep Learning AlexNet structure, Modified Crow Search (MCS) is utilized, which is mainly depends on the behavior of crow sets. The chosen Alexnet lacks in a sort of suitable supervision for upgrading execution of the proposed model that effectively aims to overfit. The proposed design, i.e., MISR, named Deep Optimal Convolutional AlexNet (DOCALN), derives the optimal values of learning rates of the ReLU activation function. Based on this optimal deep learning structure, the Low Resolution (LR) medical images can be applied. Experimentation results of our proposed model are compared with variants of Convolution Neural Networks (CNN) concerning different measures such as image quality assessment, SR efficiency analysis, and execution time.

Keywords

Deep learning super resolution optimization convolutional neural network alexnet deep learning

1 Introduction

The term Super-Resolution (SR) defines the process of re-examining at least one Low-Resolution (LR) image of a similar section on High-Resolution (HR) images [1]. It is of incredible centrality to many image processing and investigation frameworks. In any case, the SR issue is complicated because of the poorly presented condition. As it were, an LR image relates to many HR images, whereas the vast majority of them are unexpected [2, 3]. The resolutions motives for chronicle at different spatial resolution embrace capacity and broadcasting data bandwidth limitations, with better-quality Signal-To-Noise Ratio (SNR) in a few groups by more prominent bands and pixels intended for explicit determinations which don’t need high spatial resolution [4]. Numerous recreation-based SR strategies utilize at least two priors to consolidate their fundamental properties [4, 5]. The pre-trained mapping occurs between HR and LR images that are generally embraced for directing SR procedure through learning-based techniques, as per the center of learning-based strategies [5]. However, several deep learning-based techniques (DCNN, VGG net, Google net, reset) finds it useful to utilize progressive features for recreating images [5, 6].

Accurately, few research works have been carried out in single Image SR [7]. This strategy is to discover its confinements in three angles: first, it depends on the setting of little image districts; second, training combines too gradually; third, the network works for a solitary scale [8]. Using the separated features of LR, images are combined to a deep SRCNN for providing specific consideration to the image surface data, in order to uphold the general learning process progressively [9]. AlexNet, as a primary, average, fundamental, and a standout amongst the best DCNN engineering, was first proposed in our work. This method is versatile for multi-scale data of convolved feature maps in medical images and fuses the data and resolution expanded the sample Image SR by deep learning appears in Fig. 1. [10] with utilizing optimization, upgrade the learning rates of the AlexNet structure of tweaked networks to a gathering of haphazardly introduced networks. Thus, these trials are completed through a trial structure, which enables a single variable for controlling the process [11].

Fig. 1

Sample MISR Model.

DCNNs this means a network with a huge open field, characterized as the spatial degree of the input image on which the network depends to group a pixel. For SR, the advantages of the skip network while tending to its potential impediments [12]. In this procedure, one should not determine the sensor qualities like spectral response capacities. Or maybe, sensor possessions are verifiable in training medical images [13]. Here, they devised a novel Deep Convolutional network model through the optimal learning rate of the Rectified Linear Unit (ReLU) for MISR. To achieve optimum optimal values of Deep Learning AlexNet structure, Modified Crow Search (MCS) is presented. The selected Alexnet is ineffective to upgrade the model execution. The presented MISR, named Deep optimal Convolutional AlexNet (DOCALN), derives the optimal values of learning rates of the ReLU activation function. In summary, the contribution of the MISR model is talked about underneath.

We proposed a DOCALN dependent on SR for medical images, and it’s to get the most extreme resolution of images, contrasted with the existing system. Here, the improvement of SR is mainly focused on the medical images by the use of optimal AlexNet structure.

We develop the presented MISR model with four critical modules, it’s incorporate optimization system, and this model is to train the DCNN structure. The resolution of images with nonlinear focal point contortions is improved before, and after CT, MRI / ultrasound scan captures.

Optimizing the layers and learning rate of the AlexNet process, the Modified Crow Search (MCS) strategy is introduced with better probability esteems. For the test investigation, we make utilization of four regular procedure based algorithms for the assurance of our proposed structures’ accuracy.

The rest of the article organized as follows, Section 1 discussed the modern methods in MISR, and Section 2 explained briefly about recent literature. Section 3 explains proposed DOCALN-MISR, and experimental outcomes are analyzed in Section 4; at the final, conclude our work in Section 5.

2 Review of recent literature

Numerous research papers are published in the field of Image in the SR model by different researchers. Any distinguishable studies were directed in this field, as expressed in [3 –5]. In the real works of the applications, difficulties, and engineering of SR is analyzed by deep learning procedure. Various computational devices have been connected extending from straight-forward part substitution to multi-resolution examination, Bayesian inference along with variation regularization. This area examined some ongoing literature about the SR model.

2.1 Reconstruction based model

The explicit images reproduce HR images by the ray-tracing technique, and it is utilized for executing the telocentric-based light field imaging process by Shubo Zhou et al., [14], the sub-pixel moves among the precise images extricated from defocused light field information and blur in angular images. Electrical Impedance Tomography (EIT) frameworks are getting to be well known since they present a few points of interest over contending frameworks. Ricardo Augusto Borsoi et al., 2018 [15]. The re-sampling based SR technique for EIT image quality enhancement results in both engineered and also in vivo information shows that the proposed method will prompt considerable enhancements in EIT image resolution makes it increasingly aggressive by different innovations. The main obstacles recognize is that the assessment methodology generally embraced while growing new SRR procedures does not mirror the operational condition by Pawel Benecki et al. in 2018 [16]. Increasingly productive variable projection methodology rather the cyclic organize descent optimization system dependent on the regularization by Xie Qi et al., [17]. The technique consolidates enlistment, and reconstruction can conquer enrollment errors brought about by associating.

2.2 Learning-based model

An image SR strategy is utilizing dynamic generative adversarial networks (P-GANs, which can take as input with low-resolution Image and create an HR image of the preferred scaling factor. Results for Image SR demonstrate that the proposed multi organize P-GAN out-performs contending strategies and standard GANs by Dwarikanath Mahapatra et al. [18] existing CNN techniques in delving potential data in Image itself. To burrow these data, image fixes that appear to be comparative inside a similar scale and over distinctive scales are initially searched inside the input image (Jian Lu et al., 2019) [19]. The Wavelet Transformation, along with Swarm Optimization Algorithm, which gives the right SR image when contrasted with our past work, utilized a mix of Wavelet Transformation pursued by the Genetic Algorithm by Gunamani Jena et al., 2017 [20]. At that point, patch-based image features are extricated utilizing many channels and Parametric Rectified Linear Unit (PReLU) was in this way connected as the activation function (Liu, H et al., 2017) [21]. Finally, these taken out image features were utilized to remake HR images through limiting error among anticipated output image and first HR image.

The author (Chen et al., 2018) [22] proposed the Sequential Gradient Constrained Regression-based single Image SR (SGCRSR) structure that gives a successful method to consolidate the ordinary learning-based and recreation-based methodologies. At long last, a GCR-based consecutive SR system, to be specific SGCRSR, is built up for improving the nature of super-settled images bit by bit. In 2018 Honggang Chen et al., [23], the structured network is named CISRDCNN. Investigations on Image shows that proposed CISRDCNN yields state-of-the-art SR execution by frequently utilized test images as well as image sets. The consequences of CISRDCNN on original low-quality web images are exceptionally noteworthy, along with evident quality upgrades. In addition, they are investigating the use of proposed SR technique in low bit-rate image coding, prompting improved rate-distortion execution.

2.3 Recent ISR methods

SRCNN [25]: It is the primary deep Convolutional network for Image SR to acquire results. The training procedure, no soft gradient clipping, was connected for a quick training process. For the scales, it utilizes an independent model for every SR scale.

FSRCNN [26]: An improved and quickened version of the SRCNN image super-resolution is focused on ongoing applications. The training procedure, no customizable gradient, or residual learning was connected for quick training. For the scales, it shares the initial seven Convolutional layers for various SR scales.

GAN [24]: It’s utilized to get familiar with the generative model of images, which is like a specified arrangement of training images. Dynamic GAN setup where the yield of the principal stages utilizes input to the second stage and triplet loss is utilized from the second stage onwards for improving SR results.

RF [10]: This is mostly because of the utilization of group learning and sub linear search dependent on binary decision trees. RF and grouping model is to get the most exceptional learning rates of patches in a constructed tree structure by upgrading the SR procedure.

Among these applications, the medical SR model is chosen for our research work, the reason and inspiration behind the proposed work as current scenario of different infections among people, the medical scans diagnosis, treatment, and some other reason. Additionally, this SR has some problematic issues in the medical field, especially. Improving the resolutions of medical outputs, a profound learning concept is progressively useful. The resolution of images, along with nonlinear distortions, will often restrict.

3 Overview of deep learning model for ISR

Deep Learning (DL) has been generally received for some image processing applications like classification, localization, segmentation, security, and Image Super-Resolution (ISR). In our paper, we will introduce the novel DL model for ISR; this methodology simple to upgrade the resolution and nature of low to high-level Image. This improvement of images dependent on the patches and layers, fundamentally given LR Image as input, we will probably deliver HR image, which has sharp edges along with high-frequency data. Even though these methodologies have accomplished promising advancement on SISR, the overwhelming computational necessity is as yet an enormous weight even though the execution procedure. This DL having numerous methodologies for all application, our SR of image model utilized Deep Convolutional Neural Network that is DCNN, the key purpose of DCNN is learning rate and tuning parameter to improve the resolution of assumed images. The forthcoming segment explored the proposed Medical Image Super-Resolution (MISR) in depth.

3.1 Deep convolutional neural network-MISR

CNN’s are entirely supervised and need training images to increment the SR of output images. This CNN considers three essential procedures for expanding the MISR method that is medical image development, training along with testing models. The training model DCNN [27] is intended to create an association of LR as well as HR images. On the off chance that the image size (large pixel) implies the computational difficulty high. The thought behind the DCNN is applying privately trained filters to the input image and creating sub-sampled yield images ceaselessly until in-depth features are gotten. This MISR is performed by the activation function of the NN structure; it’s depicted in condition (1). This DCNN having four layers, which are the pooling layer, the Convolutional layer, the activation layer along with fully connected layer, this structure appears in Fig. 2. This function replaces the negative qualities with zeroes amid mapping. Thus, it advances in the center layers and bestows non-linearity and a little measure of calculation to accomplish a higher exactness.

Fig. 2

Structure of DCNN for MISR.

3.1.1 About DCNN Layers

Convolutional layer: Here, initialize the chosen medical images channels are connected to get the convolution of images. The noise-free medical image size is indicated as (A*B), the convoluted images weights and bias are taken by (k, l), all pixels through which the sliding happens of the walk. From the convolution, the medical image size is lessening.

Activation layer: Pooling Layer: In order to decrease the dimensionality of the feature map, we have to lessen the image processing time. The feature maps are separated from the filtered convoluted medical images. The maximum of feature map nodes inside the kernel functions to get the maximum pooling of features. It’s indicated or symbolized by $A (m) = φ (W^{T} \times m + bias)$ (1) ${CO}_{A} = max {0, bias + {\sum_{m = 0}^{n} w_{x_{1} x_{2}} \times i_{0, m 1, m 2}}}$ (2)

Both Equations (1) and (2) are the input image neurons, are weights vector, and is a nonlinear function, then is convolutional of activation value.

Fully Connected layer: The last feature map nodes are adjusted by a single feature vector and associated with the following layer’s neurons. The neuron at each layer ascertains the aggregate of weighted input sources and biased then applies the CNN activation function.

In light of the most extreme maximum feature maps and neuron, the considered input medical images are recreated to upgrade the resolution. At first, trained the DCNN [27] structure by features which we are extracted, a vigorous DCNN is proposed toward the finish of this work, which achieves superior aftereffects of the image resolution process.

3.2 Image super-resolution via AlexNet

This design is best in DCNN structure for image SR process; the principal benefit Alexnet is a superior performance with less training factors and reliable power for medical image application. This model needs to receive proper strategies for making training quicker and anticipate over-fitting because of the complex structure; it’s having numerous training parameters and also a lot of training information. For training, the structure of Alex’s net model (Fig. 3) batch normalization process and Rectified Linear Unit (ReLU) activation function utilized. It tends to improve the generalization performance of the MISR process successfully. This structure contains diverse layers it’s communicated by

C1, C2, C3, C4, and C5 = Convolutional layers

P1, P2, and P3 = Pooling layers

FC1, FC2, and FC3 = fully connected layers

Fig. 3

Structure of Alexnet model.

Usually, the Convolutional layers proceed upon the input feature maps along with sliding Convolutional parts for creating a convolved feature map. The learning feature is completed in the Convolutional layer with the two channels, which are considered independently, and which are crossed just in the third feature extraction layer for all info images. In light of the component vectors, just the image resolution is upgraded.

3.3 ReLU function

This ReLU as the activation function of the DCANet model [28] for Image SR; the output is the essential quality. This example is rehashed number of times until the Image has been consolidated spatially to a small size. ReLU is a Half-wave rectifier model that can substantially quicken the training stage, and it averts overfitting of an output image upgrade. This activation function is explained in condition (3). $A_{Re LU} (m) = {\begin{matrix} m if \\ 0 else \end{matrix} \begin{matrix} m > 0 \end{matrix}$ (3)

This function just as an activation function in each hidden layer of a NN, yet additionally as the classification capacity at the last layer of a network. Each ResBlock module just incorporates a ReLU later with the primary convolution, in addition, since the network will learn adjustments to bilinear up the sampled Image, which can be detrimental. Training as well as phenomenal restoration quality of medical images, we will improve the layers of DCNN, its coordinates by batch normalization and optimization in the CNN method, so the proposed model named as DOCALN-MISR, more insights concerning this proposed technique is described in underneath segment.

3.4 The proposed DOCALN-MISR

This area examined the DOCALN proposed model, and it’s outlined in Fig. 4. This proposed model is having four modules, which are (i) Design a DCALNM (ii) Optimal layers by MCS, (iii) Batch Normalization, and (iv) Medical Image Super-Resolution (MISR). Input produces a superior contribution for DCNN-Alexnet structure to maintain a strategic distance from noise amplification in the testing procedure. Image subtleties are too valuable to afford any losses brought about by resizing. Subsequently, the image subtleties ought to be extricated by performing a multi-scale Convolutional filter inside the network.

Fig. 4

Overall Structure of proposed MISR.

DCALNM is one of the effective techniques for image SR with a deep learning procedure. At first, the images are considered and identify the edge by utilizing canny detection [29] method, and the image perceptual loss function is reformulated as the blend of content loss and ill-disposed loss. To quicken the training procedure, and make the map learning progressively discriminative, connect a surface feature extraction layer along with the Convolutional layer.

3.5 Module I: DCALNM

This is an underlying step of the network design model; the input of the DCALNM is LR image and the output as SR of medical images. For this investigation, at first build the DCAN, it’s having numerous convolutional and feature extraction (hidden layers), from this layers, just the images are recreated, the plan of this structure clarified in following advances.

3.5.1 Input layer

The initial or first layer, let takes the LR image (m), and extract the features that are (68 maps) from our corner, and each Image with ten patches is chosen for contributing to the training network, which is utilized as info images Center area. The feature maps are produced by Convolutional weights, and bias esteems with extracted highlights it’s communicated by $F_{SR} (m, w_{cw}, a_{cw - 1}) = max (w_{cw} \times m + a_{cw - 1}, 0)$ (4)

The input and desired output of DCALN are comparable, so learning residual Image is more suitable. Hence, the model adopts the residual learning strategy by this module.

3.5.2 Convolutional (C1-C5)

This layer is essential for the proposed model, it has pre-trained filters to apply the LR images, and here five layers region used to locate the convoluted procedure. The size of an open field of the neuron is controlled by the size of the Convolutional kernel, the optimal size of the kernel is to extricate compelling neighborhood features in the scope of kernel along with portrayal capacity. This convolution procedure, because of activation value (ReLU), for optimizing the learning rate of proposed model capacity portrayed in the forthcoming area.

3.5.3 Pooling (P1, P2, and P3)

Here, the Generated feature map is abridged by pooling administrators; for example, most extreme, average pooling, it has three-layer work that is maximum, average and minimum pooling values. This esteem determined by utilizing the learning rate of network structure. The hidden layer of learning along these lines can’t depend on the nearness of specific different features of the P1, P2, and P3 layer, that makes learning features will prompt more grounded strength.

3.5.4 Last layers (FC1, FC2, and FC3)

This a last layer of DCALN that makes the learning features of the two channels cross-blending to acquire feature vector. Neurons at each layer ascertain the sum of weighted inputs and biased then apply optimal learning esteems. From the optimal activation function values, it lessens the top-1 and top-5 are error rates as contrasted with the non-covering plan, which produces an output of proportionate elements of SR images. The Equ for output layer calculation is $O (sr) = \frac{e^{{fc}_{i} (sr)}}{\sum_{i = 1} {fc}_{i - 1}} (i)$ (5)

To enhance the robustness in the FC1, FC2, and FC3, employed recently-developed regularization method, it is called as Dropout. Thus the learning of hidden layers is not dependent on features of pooling layers.

3.6 Module II: Layer optimization

MISR process optimizes the layers and learning rate of DCALNM by utilizing MCS. It’s a meta-heuristic algorithm that is created by using Askarzadeh, and it is enlivened by the intellect behavior of the crows [30]. This DCNN model, the parameters updated by using Equation (7). This optimal solution attaining process having three essential steps, which are,

They live in the form of a flock

Memorize food hiding places

Probability evaluation and protect their caches from being pilfered

In our MISR model, the weights are represented in Equation (6), for pre-trained AlexNet architecture, and the relationships among the weight parameters. These weights are convolved retorts on the previous feature map and the pooling layer function of MISR. $\begin{matrix} W^{\circ} = A_{f} (L^{\circ}) & W^{\circ} \\ L^{\circ} = Y^{\circ} \times W^{\circ - 1} \end{matrix}$ (6)

The DCNN [27] structure along the diagonal direction as to streamlining the learning estimations of the MISR process. Optimal Learning rate and layers of network structure get the most extreme precision of the SR process. The fine-tuned networks beat scratch statement. Thus the worldwide learning rate played out the best qualities. Whereas the target capacity of MISR appears in (7),

$\begin{matrix} O_{L} R = {{{∥ Y^{e} ∥}^{2} + P (Y, y^{e})} \\ + {\sum_{0 - 1}^{0 - 1} S_{\circ} {[∥ y^{\circ} ∥]}^{2} - rand}} \end{matrix}$ (7)

Here {∥ Y^e ∥ ² + P (Y, y^e) } are e margin and squared hinge loss of network classifiers. The pre-trained DOCALN architecture not only acquires Convolutional kernels, and also in each hidden layer, it directly makes a good label prediction and also gives a strong push to have sensible and discriminative features in each layer.

3.7 Create the New position of DOCALN

Introduce the crow “r” realize the crows to another position of the search space. If the new position of the crow is achievable, it invigorates its position. Approximately it receives other than what’s expected; the crow stays in the present position, and it does not move to deliver in a new position. That is

$\begin{matrix} Y^{o, i + 1} = \\ {\begin{matrix} y^{o, i} + {rand}_{i} \times q_{lengath}^{o, i} (o_{n}^{n, i} - m^{o, i}) \\ Randomcrowposition \end{matrix} if \begin{matrix} {rand}^{n} \geq Pr o^{n, i} \\ otherwise \end{matrix} \end{matrix}$ (8)

$\begin{matrix} {New}_{m} {emory}^{i, i + 1} = \\ {\begin{matrix} O_{L} R^{i} \\ {memory}^{i, iter} \end{matrix} \begin{matrix} f ({LR}^{i, i + 1}) > ({memory}^{i}) \\ otherwise \end{matrix} \end{matrix}$ (9)

The above conditions (8 and 9) are utilized to locate the optimal crow position of DOCALN, constrained by the probability values. The crow remains in the present position, and it does not move to a new position created. Sought after, fitness value for the revived position is settled. In condition (10), the probability calculation dependent on the modified procedure that crows’ random position can be characterized as: $Pr o = 0.781 [1 - \frac{LRofi}{{LR}_{max}}]$ (10)

This calculation is iteratively associated with different iterations intending to meet a satisfactory solution. It justifies referencing that the discretionary position of the crow is confined by the investigation rate at the present cycle. Given the optimal qualities, the error function is determined, it’s expressed by loss function that is $Loss \Rightarrow {MSR}_{error} = 1 / 2 {[a_{msr} - o_{msr}]}^{2}$ (11)

The loss function of the feature extraction layer is planned to enforce texture features of the restored Image that will be consistent with accurate ground images. From this Equ to identify the quality of MISR work in DOCALN.

3.8 Module III: Batch normalization

From the optimal learning rate and layers in DOCALN is high exactness and execution, and keeps the training from the non-linearity esteems. Two additional parameters are included per activation, and also parameters will be learned in the network training stage, and this procedure appears in Fig. 5. On the off chance that initiations that rely upon the mini-batch permit practical training yet are neither vital nor attractive amid derivation, we need the output to depend just on the info image it shows in condition (12). $N_{b} = Optimal_Re LU (c_{1 . . n} \times {fc}_{1 . . n} \times m + b)$ (12)

Fig. 5

Optimal ReLU based batch normalization.

This condition clarifies the optimal activation function and learning rates of Convolutional and fully connected layers, and then the bias is to discover the normalization esteem. From this procedure, the training image is reproduced to upgrade the resolution of medical images. This optimal Alexnet is anything but difficult to upgrade the quality and expanding the SR by standardized learning rates.

3.9 Module IV: Optimal deep learning-based MISR

This is the last module of our proposed MISR process; in light of this, the training system, the objective of every module will be accomplished, whereas the last joint optimization strategy limits prediction error. Besides, straight-forward training a deep network is troublesome. Introducing a deep network with educated parameters is advantageous for acquiring a steady training method and also with a quick convergence rate. The last SR is communicated by ${SR}_{m} = \frac{1}{2 K} \sum_{i = 1}^{K} ∥ (M_{i}^{train}, Φ_{opt}) - T_{i}^{image} ∥$ (13)

This DOCALN gives specific consideration to the image texture data, instead of authorizing the general learning method increasingly successfully. The effectiveness of straight-forward deeper structures for an SR is apparent for SR applications for image characterization errands. We assess the adequacy of the depth of the network as for reconstruction execution of medical images.

4 Experimental results analysis

In this result analysis section, the implementation settings and the effectiveness of our proposed DOCANM-MISR for training and testing medical images results are described in detail [31]. The considerable medical images are LR, and we will analyze results with K- fold validation. Moreover, this innovative MISR implemented in MATLAB 2018, i3 processor, and 4GB RAM; this performance analysis is entirely based on different quality measures. For experimentation, the parameter used for this method is batch size: 8, learning rate: 0.02, step size or epoch: 10000, score threshold value: 0.7, minimum dimension: 600, and the maximum dimension: 1024.

Figure 6 clarifies the aftereffects of medical images that we are taken for the examination. Figure (a) explains the original images of the brain, ECG, retina, lung, liver, and kidney. After the application of edge smoothening technique, we attained the images with smoothened edges; it is clearly shown in Figure (b). Next to that, the low-resolution images of each input image are depicted in Figure (c). Figure (d), (e), (f), (g) explains the imaging analysis of existing techniques like SR- CISRDCNN, SRCNN, FSRCNN, DCALNM, respectively; where the abnormal portions are identified and marked as a rectangular green box. The proposed SR images- DOCALN results are figured in section (h).

Fig. 6

(a) Original images, (b) Edge Smoothing, (c) LR images and (d) SR- CISRDCNN, (e) SRCNN, (f) FSRCNN, (g) DCALNM and (h)SR images- DOCALN (Our method).

Tables 1, 2, and 3 demonstrate that the peak signal to noise ratio (PSNR), information fidelity criterion (IFC) and structural similarity (SSIM) of the DOCALN over the compared algorithms. The images which we are taken for the analysis are the brain, ECG, Retina, Lung Nodule, liver, and kidney. Otherwise, the results may or may not be observable in various cases. The highest average based LR value for PSNR analysis is 31.27333, which is attained at 0.5 optimal LR. Similarly, the highest average based LR value for SSIM is 0.93, which is attained at 0.5 optimal LR. In addition to this, the highest average based LR value for IFC is 2.49, which is attained at 0.5 optimal LR. Finally, it examines that DOCALN attains precise perceptual quality perfection over original LR images.

Table 1

PSNR (dB) of DOCALN via LR

Optimal LR	Brian	ECG	Retinal image	Lung nodule	Liver	Kidney images	Average based LR
0.2	23.22	22.45	24.85	23.28	26.75	28.52	24.845
0.3	29.45	22.78	26.28	24.722	27.85	29.47	26.75867
0.45	29	29.78	30.78	24.89	29.44	30	28.98167
0.5	31.22	33.22	31.22	29.45	32.45	30.08	31.27333
Average	28.2225	27.0575	28.2825	25.5855	29.1225	29.5175	–

Table 2

SSIM of DOCALN via LR

Optimal LR	Brian	ECG	Retinal image	Lung nodule	Liver	Kidney images	Average based LR
0.2	0.75	0.80	0.78	0.62	0.82	0.92	0.78
0.3	0.86	0.78	0.82	0.68	0.86	0.94	0.82
0.45	0.88	0.89	0.86	0.79	0.88	0.88	0.86
0.5	0.895	0.933	0.955	0.89	0.96	0.94	0.93
Average	0.84625	0.85075	0.85375	0.745	0.88	0.92	–

Table 3

IFC of DOCALN via LR

Optimal LR	Brian	ECG	Retinal image	Lung nodule	Liver	Kidney images	Average based LR
0.2	1.274	2.25	1.52	0.89	0.86	1.71	1.42
0.3	1.33	0.968	1.45	1.85	0.899	1.84	1.39
0.45	1.41	0.89	1.85	2.22	1.85	2.25	1.75
0.5	1.88	2.2	2.28	2.89	2.74	2.96	2.49
Average	1.4735	1.577	1.775	1.9625	1.587	2.19	–

Figure 7 illustrates the performance measures of the proposed model for six different medical images. The images which we are taken for the analysis are

Fig. 7

Sen, Spc, AUC analysis.

The brain, ECG, Retina, Lung Nodule, liver, and kidney. The maximum accuracy is achieved in the ECG image analysis when compared to other images. The detection accuracy is improved in the DOCALN because of its enhanced process in the DCNN based optimization. This metaheuristic algorithm optimizes the layers and learning rate of DCALNM. Hence it gives better performance in terms of accuracy, specificity, and sensitivity.

The relation between feature maps and PSNR value is analyzed and compared among different MISR images (SRCNN, CNN, DCALNM, FSRCNN, and the proposed DOCALN) in Fig. 8. At 10 mapping feature, the range of PSNR value is 18 to 23, and it gradually increases for the higher mapping values of feature vector (the range is up to 30 dB). From the graphical analysis, the proposed DOCALN (MISR) images attain high PSNR value compared to other images.

4.1 Comparative analysis

In order to compare the perceptual quality of Test Images (TIs), the model considers the analysis of K-fold validation for the values K = 5, K = 10, and K = 15. The results of TI1 (Brain), TI2 (ECG), TI3 (Retina), TI4 (Lung Nodule), TI5 (liver), and TI6 (kidney) are presented in Table 4. From the quality investigation, we can reveal that the DOCALN (MISR) image creates fewer antiques and better preserves the principle structures that are to recognize the abnormal portions of medical images; in specific nodules in lung image (explained in Fig. 6 (h)). For each k-fold validation, PSNR, SSIM, and IFC are examined, and the result demonstrates that the highest PSNR, IFC, and SSIM are accomplished in 15 K-fold.

Table 4
Quality Results for Test Image via K fold validation

Images K = 5 K = 10 K = 15

TI1 PSNR SSIM IFC PSNR SSIM IFC PSNR SSIM IFC

TI2 22.22 0.85 1.52 23.45 0.79 1.72 22.85 0.66 1.59

TI3 23.41 0.79 1.53 26.41 0.82 1.76 28.8 0.74 1.99

TI4 21.45 0.81 1.72 26.45 0.834 1.82 28.99 0.76 2.04

TI5 28.22 0.76 1.59 31.22 0.864 1.78 30.45 0.85 2.12

TI6 28.74 0.78 1.60 30.71 0.91 1.92 31.07 0.92 2.19

Images	K = 5	K = 10	K = 15
TI2	22.22	0.85	1.52	23.45	0.79	1.72	22.85	0.66	1.59
TI3	23.41	0.79	1.53	26.41	0.82	1.76	28.8	0.74	1.99
TI4	21.45	0.81	1.72	26.45	0.834	1.82	28.99	0.76	2.04
TI5	28.22	0.76	1.59	31.22	0.864	1.78	30.45	0.85	2.12
TI6	28.74	0.78	1.60	30.71	0.91	1.92	31.07	0.92	2.19

The comparative analysis of different imaging techniques (SRCNN, CNN, DCALNM, FSRCNN, and the proposed DOCALN) is presented in Fig. 9 by evaluating the performance measures PSNR in Fig. 9 (a), SSIM in Fig. 9 (b) and IFC in Fig. 9 (c). The distributions of PSNR/IFC/SSIM gains of MI1, MI2, MI3, MI4, MI5, and MI6 are plotted in the bar graph. The results demonstrate the robustness as well as the stability of the proposed DOCALN model. Overall, the SR-DOCALN executes better than JPEG at low bit-rates employing both objective and subjective evaluations.

Fig. 8

The relation between feature maps & PSNR.

Fig. 9

(a) PSNR (b) SSIM and (c) IFC.

The processing time analysis of different existing (SRCNN, FSRCNN, CNN, and DCALNM) and proposed DOCALN methods are discussed in Table 5. For structure generation, SRCNN is 45, FSRCNN is 59, CNN is 62, DCALNM is 66, and DOCALN is 68. For feature vector mapping, SRCNN is 66, FSRCNN is 75, CNN is 75, DCALNM is 88, and DOCALN is 61; in this proposed one takes minimum time. Here, in our work, only we use optimization to select the optimal features; totally, it needs 234 seconds to detect the abnormal portion of medical images. The proposed DOCALN method achieves the most optimal values to diagnose diseases since it takes more time compared to other techniques.

Table 5

CPU Time (sec) analysis

Methods	Structure gen	Feature maps	Optimization	Total (sec)
SRCNN	45	66	–	111
FSRCNN	59	75	–	134
CNN	62	75	–	137
DCALN	66	88	–	154
DOCALN	68	61	105	234

The efficiency analysis of medical super-resolution images is depicted in Fig. 10 as a bar graph. The efficiency attained for the existing technique is SRCNN achieves 88.85%, FSRCNN attains 89.55%, CNN achieves 75.55%, DCALN gives 86.5%, and the proposed DOCALN gives 96.45%. By analyzing the PSNR, IFC, and SSIM achieve over state-of-the-art methods that demonstrate efficiency as well as the superiority of DOCALN. In summary, SR-DOCALN reaches state-of-the-art performance among different existing approaches.

Fig. 10

SR Efficiency analysis.

5 Conclusion

In this paper, we proposed an innovative MISR algorithm; it shows the excellent SR accuracy. The designed DOCALN of medical image quality measures compared with other deep learning systems. As a result, a conventional DCNN was trained to manage various scenarios of distortions as well as different SR scales. Be that as it may, because of the lack of high-performance computing devices, and the high complexity of the training process, the parameters of the proposed framework MCS optimization is used. The benefit of this optimization in the proposed detection technique is to get optimal values of learning rates of the ReLU activation function. Based on this optimal deep learning structure, the LR is enhancing the resolution of medical images. This LR images which will draw in more researchers to concern these issues. Our proposed technique can accomplish the quality outcomes by achieving 30.12 of PSNR, 0.88 of SSIM, and 1.86 of IFC when compared to the traditional deep learning techniques and some different literary works. In addition, the proposed DOCALN gives a maximum SR efficiency of 96.45%. In future decades, the same DOCALN is executed for a video frame and other surveillance applications by varying the parameter setting of deep learning structure.

References

Feng

, Lai

, Xie

and Zhu

, Image super-resolution via a densely connected recursive network, Neurocomputing 316 (2018), 270–276.

Tang

, Luo

, Peng

and Li

, A joint residual network with paired ReLUs activation for image super-resolution, Neurocomputing 273 (2018), 37–46.

Zhu

, Zhan

and Zhang

, Stacked U-shape Networks with Channel-wise Attention for Image Super-Resolution, Neurocomputing (2019).

Chang

, Hung

K.W.

and Jiang

, Deep learning-based image Super-resolution for nonlinear lens distortions, Neurocomputing 275 (2018), 969–982.

Nayak

and Patra , New single-image super-resolution reconstruction using MRF model, Neurocomputing 293 (2018), 108–129.

Becherer

, Pecarina

, Nykl

and Hopkinson

, Improving optimization of convolutional neural networks through parameter fine-tuning, Neural Computing and Applications (2017), 1–11.

Lanaras

, Bioucas-Dias

, Galliani

, Baltsavias

and Schindler

, Super-resolution of Sentinel-2 images: Learning a globally applicable deep neural network, ISPRS Journal of Photogrammetry and Remote Sensing 146 (2018), 305–319.

Garcia-Garcia

, Orts-Escolano

, Oprea

, Villena-Martinez

and Garcia-Rodriguez

, A review on deep learning techniques applied to semantic segmentation. arXiv preprint arXiv:1704.06857. (2017).

Moustafa

, Ebied

H.M.

, Helmy

, Nazamy

T.M.

and Tolba

M.F.

, Optimization Methods for Medical Image Super-Resolution Reconstruction, In Applications of Intelligent Optimization in Biology and Medicine, Springer, (2016), pp. 135–157.

10.

Wei

, Zhou

, Wu

, Pu

, Wang

and Yang

, Medical image super-resolution by using multi-dictionary and random forest, Sustainable Cities and Society 37 (2018), 358–370.

11.

Han

, Zhong

, Cao

and Zhang

, Pre-trained alexnet architecture with pyramid pooling and supervision for high spatial resolution remote sensing image scene classification, Remote Sensing 9(8) (2017), 848.

12.

, Ren

, Yang

, Wan

and Wang

, Texture-centralized deep convolutional neural network for single Image super-resolution, In 2017 Chinese Automation Congress (CAC) IEEE, 3707–3710.

13.

Sinha

, Verma

and Haidar

, Optimization of convolutional neural network parameters for image classification, In 2017 IEEE Symposium Series on Computational Intelligence (SSCI), 1–7. IEEE.

14.

Zhou

, Yuan

, Su

, Ding

and Wang

, Multiframe super-resolution reconstruction method based on light field angular images, Optics Communications 404 (2017), 189–195.

15.

Borsoi

R.A.

, Aya

J.C.C.

, Costa

G.H.

and Bermudez

J.C.M.

, Super-resolution reconstruction of electrical impedance tomography images, Computers & Electrical Engineering 69 (2018), 1–13.

16.

Benecki

, Kawulok

, Kostrzewa

and Skonieczny

, Evaluating super-resolution reconstruction of satellite images, Acta Astronautica 153 (2018), 15–25.

17.

and Yi

C.W.

, Super-resolution Reconstruction Base on a New Image Registration, In 2014 International Symposium on Computer, Consumer and Control, IEEE, 318–321.

18.

Mahapatra

, Bozorgtabar

and Garnavi

, Image super-resolution using progressive generative adversarial networks for medical image analysis, Computerized Medical Imaging and Graphics 71 (2019), 30–39.

19.

, Hu

and Sun

, A deep learning method for image super-resolution based on geometric similarity, Signal Processing: Image Communication 70 (2019), 210–219.

20.

Jena

, Panda

S.S.

, Rajesh

B.V.

and Jena

, Image Super-Resolution Using Wavelet Transformation and Swarm Optimization Algorithm, In International Conference on Intelligent Human Systems Integration, Springer, Cham, (2018), 444–449.

21.

Liu

, Xu

, Wu

, Guo

, Ibragimov

and Xing

, Learning deconvolutional deep neural network for high-resolution medical image reconstruction, Information Sciences 468 (2018), 142–154.

22.

Chen

, He

, Qing

, Teng

and Ren

, SGCRSR: Sequential gradient constrained regression for single image super-resolution, Signal Processing: Image Communication 66 (2018), 1–18.

23.

Chen

, He

, Ren

, Qing

and Teng

, CISRDCNN: Super-resolution of compressed images using deep convolutional neural networks, Neurocomputing 285 (2018), 204–219.

24.

, Walia

and Babyn

, Generative adversarial network in medical imaging: A review, arXiv preprint arXiv:1809.07294, (2018).

25.

Kim

, Kwon Lee

and Mu Lee

, Accurate image super-resolution using very deep convolutional networks, In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2016), 1646–1654. 686

26.

Dong

, Loy

C.C.

and Tang

, Accelerating the super-resolution convolutional neural network, In European conference on computer vision, Springer, Cham, (2016), 391–407.

27.

Wang

, Jiao

, Yin

, Zhao

, Han

and Sun

, Underwater sonar image classification using adaptive weights convolutional neural network, Applied Acoustics 146 (2019), 145–154.

28.

Unnikrishnan

, Sowmya

and Soman

K.P.

, Deep AlexNet with Reduced Number of Trainable Parameters for Satellite Image Classification, Procedia Computer Science 143 (2018), 931–938.

29.

Gaurav

and Ghanekar

, Image steganography based on Canny edge detection, dilation operator and hybrid coding, Journal of Information Security and Applications 41 (2018), 41–51.

30.

Javidi

, Salajegheh

and Salajegheh

, Enhanced crow search algorithm for optimum design of structures, Applied Soft Computing (2019). https://github.com/sfikas/medical-imaging-datasets.