Обработка на изображения (PDF)

1. Въведение Обработка на изображения, магистри, ФМИ Обработка на изображения в роботиката Обработка на изображения в роботиката Обработка на изображения в роботиката Обработка на изображения Цифрови изображения Цифрови изображения Цифрови изображения Дължини на вълни Дължини на вълни Дъл...

1. Въведение Обработка на изображения, магистри, ФМИ Обработка на изображения в роботиката Обработка на изображения в роботиката Обработка на изображения в роботиката Обработка на изображения Цифрови изображения Цифрови изображения Цифрови изображения Дължини на вълни Дължини на вълни Дължини на вълни Обработка на изображения за инспекция Други приложения Ултразвук Примерна система за обработка на изображения Оптически илюзии Оптически илюзии Оптически илюзии Примери от практиката https://youtu.be/0_13uh4YsJM https://youtu.be/6xu-E_w8CoM https://youtu.be/wg8YYuLLoM0 Научно-изследователски примери https://youtu.be/Z9KSxnWtZj8 https://youtu.be/pT_JIo4Qe5s https://youtu.be/AeIAdpC0O2k 2. Основни понятия и дефиниции Обработка на изображения, магистри, ФМИ Представяне на цифрови изображения Ако резолюцията на цифровото изображение е М x N пиксела, то ще го представяме по следния начин: или като матрица: Представяне на цифрови изображения Всеки елемент на матрицата ще наричаме пиксел. Нивата на интензитет са L на брой и съответно всеки пиксел има стойност в интервала: [0, L - 1]. Обикновено L е степен на двойката. Съседи на пиксел Нека p e с координати (x, y). Тогава p има: - 4 хоризонтални и вертикални съседни пиксела: - 4 диагонални съседни пиксела: Общо означение:  Съседство Ако V е множеството на нивата на интензитет, чрез коите ще се дефинира съседност на пиксели. В двоично изображение V={1}, ако за съседни считаме всички пиксели, които стойност 1. В полутоново изображение е възможно V да съдържа повече елементи. Ако нивата са от 0 до 255, то V е произволно подмножество на тези 256 стойности.  Съседство Ще разглеждаме следните три типа съседност: - 4-съседност: Два пиксела p и q със стойности от V са 4-съседни, ако q e в множеството N4( p). - 8-съседност: Два пиксела p и q със стойности от V са 8-съседни, ако q e в множеството N8( p). - m-съседност: Два пиксела p и q със стойности от V са m-съседни, ако: - q e в множеството N4( p) или - q e в множеството ND( p) и в сечението на N4( p) и N4(q) няма пиксели със стойности от V. Съседство  Път Път ( или крива) от пиксел с координати (x, y) до пиксел с координати (s, t) ще наричаме последователността от различни пиксели с координати: (x0, y0),(x1, y1),...,(xn, yn), където (x0, y0) = (x, y), (xn, yn) = (s, t) и пикселите (xi, yi) и (xi-1, yi-1) са съседни за 1 results in spike at ends of histogram ⚫ Spikes and Gaps in manipulated images (not original) Image Defects: Effect of Image Compression ⚫ Histograms show impact of image compression ⚫ GIF compression reduces the dynamic range to only few intensities (quantization) Original Histogram Original Histogram after GIF conversion Image Fix? Scaling image by 50% and Interpolating values recreates some lost colors But GIF artifacts still visible Effect of Image Compression ⚫ Example: Effect of JPEG compression over line graphics ⚫ JPEG compression designed for color images Original histogram has only 2 intensities (gray and white) JPEG image appears dirty, fuzzy and blurred Its Histogram contains gray values not in original Large Histograms: Binning ⚫ High resolution image can yield very large histogram ⚫ Example: 32‐bit image = 232 = 4,294,967,296 columns ⚫ Such a large histogram impractical to display ⚫ Solution? Binning! ⚫ Combine ranges of intensity values into histogram columns Number (size of set) of pixels Pixel’s intensity is such that between ai and ai+1 Calculating Bin Size ⚫ Typically use equal sized bins ⚫ Bin size: ⚫ Example: To create 256 bins from 14‐bit image  Binned Histogram ⚫  Color Image Histograms Two types: 1. Intensity histogram: ⚫ Convert color image to gray scale ⚫ Display histogram of gray scale 2. Individual Color Channel Histograms: 3 histograms (R,G,B) Color Image Histograms ⚫ Both types of histograms provide useful information about lighting, contrast, dynamic range and saturation effects ⚫ No information about the actual color distribution ⚫ Images with totally different RGB colors can have same R, G and B histograms  Cumulative Histogram ⚫ Useful for certain operations, e.g. histogram equalization ⚫ Analogous to the Cumulative Density Function (CDF) ⚫ Definition: ⚫ Recursive definition ⚫ Monotonically increasing Last entry of Total number of Cum. histogram pixels in image Point Operations ⚫ Point operations changes a pixel’s intensity value according to some function (don’t care about pixel’s neighbor) ⚫ Also called a homogeneous operation ⚫ New pixel intensity depends on ⚫ Pixel’s previous intensity I(u,v) ⚫ Mapping function f ⚫ Does not depend on ⚫ Pixel’s location (u,v) ⚫ Intensities of neighboring pixels Homogeneous Point Operations ⚫ Addition (Changes brightness) ⚫ Multiplication (Stretches/shrinks image contrast range) ⚫ Real‐valued functions ⚫ Quantizing pixel values ⚫ Global thresholding ⚫ Gamma correction Point Operation Pseudocode ⚫ Input: Image with pixel intensities I(u,v) defined on [1.. w] x [1.. H] ⚫ Output: Image with pixel intensities I’(u,v) for v = 1.. h for u = 1.. w set I(u, v) = f (I(u,v)) Non‐Homogeneous Point Operation ⚫ New pixel value depends on: ⚫ Old value + pixel’s location (u,v) Clamping ⚫ Deals with pixel values outside displayable range ⚫ If (a > 255) a = 255; ⚫ If (a < 0) a = 0; ⚫ Function below will clamp (force) all values to fall within range [a,b] Example: Modify Intensity and Clamp ⚫ Point operation: increase image contrast by 50% then clamp values above 255  Inverting Images (Image Negatives) ⚫ 2 steps 1. Multiple intensity by ‐1 2. Add constant (e.g. amax) to put result in range [0,amax] Original Inverted Image  Inverting Images (Image Negatives) ⚫Image negatives useful for enhancing white or Images taken from Gonzalez & Woods, Digital Image Processing (2002) grey detail embedded in dark regions of an image ⚫ Note how much clearer the tissue is in the negative image of the mammogram below Original Negative s = 1.0 - r Image Image Thresholding ⚫ Thresholding Example Thresholding and Histograms ⚫ Example with ath = 128 ⚫ Thresholding splits histogram, merges halves into a0 a1  Basic Grey Level Transformations Images taken from Gonzalez & Woods, Digital Image Processing (2002) ⚫ 3 most common gray level transformation: ⚫ Linear ⚫ Negative/Identity ⚫ Logarithmic ⚫ Log/Inverse log ⚫ Power law ⚫ nth power/nth root  Logarithmic Transformations ⚫Maps narrow range of input levels => wider range of output values ⚫ Inverse log transformation does opposite transformation ⚫ The general form of the log transformation is New pixel value s = c * log(1 + r) Old pixel value ⚫ Log transformation of Fourier transform shows more detail s = log(1 + r)  Power Law Transformations ⚫ Power law transformations have the form Images taken from Gonzalez & Woods, Digital Image Processing (2002) Power s=c*r γ Old pixel value New pixel value Constant ⚫Map narrow range of dark input values into wider range of output values or vice versa ⚫Varying γ gives a whole family of curves  Power Law Example Original Images taken from Gonzalez & Woods, Digital Image Processing (2002) ⚫Magnetic Resonance (MR) image of fractured s = r 0.6 human spine s = r 0.4 ⚫Different power values highlight different details Intensity Windowing ⚫ A clamp operation, then linearly stretching image intensities to fill possible range ⚫ To window an image in [a,b] with max intensity M Intensity Windowing Example Contrasts easier to see Bit-plane Slicing Bit-plane Slicing 10/10/2024 48 Bit-plane Slicing Point Operations and Histograms ⚫ Effect of some point operations easier to observe on histograms ⚫ Increasing brightness ⚫ Raising contrast ⚫ Inverting image ⚫ Point operations only shift, merge histogram entries ⚫ Operations that merge histogram bins are irreversible Combining histogram operation easier to observe on histogram Automatic Contrast Adjustment ⚫ Original intensity range If amin = 0 and amax = 255 New intensity range Effects of Automatic Contrast Adjustment Linearly stretching range causes gaps in histogram Original Result of automatic Contrast Adjustment Modified Contrast Adjustment ⚫ Histogram Equalization ⚫ Adjust 2 different images to make their histograms (intensity distributions) similar ⚫ Apply a point operation that changes histogram of modified image into uniform distribution Histogram Cumulative Histogram Histogram Equalization Spreading out the frequencies in an image (or equalizing the image) is a simple way to improve dark or washed out images Can be expressed as a transformation of histogram ⚫ rk: input intensity ⚫ sk: processed intensity ⚫ k: the intensity range (e.g 0.0 – 1.0) processed intensity sk = T (rk ) input intensity Intensity range (e.g 0 – 255) Images taken from Gonzalez & Woods, Digital Image Processing (2002) Equalization Transformation Function  Equalization Transformation Functions Images taken from Gonzalez & Woods, Digital Image Processing (2002) Different equalization function (1‐4) may be used Images taken from Gonzalez & Woods, Digital Image Processing (2002) Equalization Examples 1 Images taken from Gonzalez & Woods, Digital Image Processing (2002) Equalization Examples 2 Images taken from Gonzalez & Woods, Digital Image Processing (2002) Equalization Examples 4 3  Linear Histogram Equalization ⚫ Histogram cannot be made exactly flat – peaks cannot be increased or decreased by point operations. ⚫ Following point operation makes histogram as flat as possible: (assuming M x N image and pixels in range [0, K ‐ 1]) Point operation that returns Linear equalized value of a Cumulative Histogram: Σ how many times intensity a occurs  Effects of Linear Histogram Equalization Original Image Image after Linear Equalization Original Histogram after histogram Linear Equalization Cumulative Histogram Cumulative After Linear Histogram Equalization Images and Probability  Histogram Specification ⚫ Real images never show uniform distribution (unnatural) ⚫ Most real images, distribution of pixel intensities is gaussian ⚫ Histogram specification ⚫ modifies an image’s histogram into an arbitrary intensity distribution (may not be uniform) ⚫ Image 1’s histogram can also be used as target for image 2 ⚫ Why? Makes images taken by 2 different cameras to appear as if taken by same camera  Histogram Specification ⚫ Find a mapping such that distribution of a matches some reference distribution.i.e Mapping function: maps distribution on right to equivalent point (same height) On distribution on left to convert original image IA into IA’ such that i.e. a and a’ have same height (b) on different CDF distributions PA(a) P-1 R(b) Adjusting Linear Distribution Piecewise ⚫ In practice, reference distribution may be specified as a piecewise linear function ⚫ 2 endpoints are fixed Adjusting Linear Distribution Piecewise For each segment, linearly Interpolate to find any value We also need the inverse mapping Adjusting Linear Distribution Piecewise Adjusting Linear Histogram Piecewise  Histogram Matching ⚫ Prior method needed reference distribution to be invertible Has to be invertible ⚫ What if reference histogram is not invertible? ⚫ For example not invertible if histogram has some intensities that occur with probability 0? i.e. p(k) = 0 ⚫ Use different method called histogram matching Histogram Matching Makes the intensity profiles of two images IA and IB look as similar as possible By “matching” their cumulative histograms. Works fine for images with similar content.  Adjusting to a Given Histogram Reference Height of original Cumulative Intensity on Distribution histogram Matched intensity Original intensity Adjusting to a Given Histogram  Adjusting to a Given Histogram Original histogram after matching original histogram CDF of original CDF of original histogram after histogram matching Adjusting to a Given Histogram  Local Histogram Processing Define a neighborhood and move its center from pixel to pixel At each location, the histogram of the points in the neighborhood is computed. Either histogram equalization or histogram specification transformation function is obtained Map the intensity of the pixel centered in the neighborhood Move to the next location and repeat the procedure Local Histogram Processing: Example  Using Histogram Statistics for Image Enhancement Average Intensity L −1 M −1 N −1 1 m =  ri p (ri ) = MN   f ( x, y ) x =0 y =0 i =0 L −1 un (r ) =  (ri − m) n p(ri ) i =0 Variance L −1 M −1 N −1 1  = u2 (r ) =  (ri − m) p(ri ) =    f ( x, y ) − m  2 2 2 i =0 MN x =0 y =0 Using Histogram Statistics for Image Enhancement: Example 79 Adaptive histogram equalization (AHE) computes several histograms corresponding to a distinct section of the image redistribute the lightness values of the image suitable for improving the local contrast and enhancing the definitions of edges in each region of an image Properties of AHE The size of the neighborhood region is a parameter of the method. It constitutes a characteristic length scale: contrast at smaller scales is enhanced, while contrast at larger scales is reduced. Due to the nature of histogram equalization, the result value of a pixel under AHE is proportional to its rank among the pixels in its neighborhood. This allows an efficient implementation on specialist hardware that can compare the center pixel with all other pixels in the neighborhood. Properties of AHE When the image region containing a pixel's neighbourhood is fairly homogeneous regarding to intensities, its histogram will be strongly peaked, and the transformation function will map a narrow range of pixel values to the whole range of the result image. This causes AHE to overamplify small amounts of noise in largely homogeneous regions of the image. Properties of AHE AHE overamplifies noise in relatively homogeneous regions of an image contrast limited adaptive histogram equalization (CLAHE) prevents this by limiting the amplification CLAHE limits the amplification by clipping the histogram at a predefined value before computing the CDF This limits the slope of the CDF and therefore of the transformation function It is advantageous not to discard the part of the histogram that exceeds the clip limit but to redistribute it equally among all histogram bins. Contrast Limited AHE The redistribution will push some bins over the clip limit again (region shaded green in the figure), resulting in an effective clip limit that is larger than the prescribed limit and the exact value of which depends on the image. If this is undesirable, the redistribution procedure can be repeated recursively until the excess is negligible. Efficient computation by interpolation AHE, both with and without contrast limiting, requires the computation of: a different neighborhood histogram; transformation function for each pixel in the image. This makes the method very expensive computationally. Efficient computation by interpolation Interpolation allows a significant improvement in efficiency without compromising the quality. The image is partitioned into equally sized rectangular tiles (tiles with 8 columns and 8 rows) Histogram, CDF and transformation function is then computed for each tile. The transformation functions are appropriate for the tile center pixels. All other pixels are transformed with up to four transformation functions of the tiles with center pixels closest to them and are assigned interpolated values. Efficient computation by interpolation Blue pixels are bilinearly interpolated, Green pixels close to the boundary are linearly interpolated, Red corner pixels are transformed with the transformation function of the corner tile The interpolation coefficients reflect the location of pixels between the closest tile center pixels, so that the result is continuous as the pixel approaches a tile center. Efficient computation by incremental update of histogram An alternative to tiling the image is to "slide" the rectangle one pixel at a time and only incrementally update the histogram for each pixel, by adding the new pixel row and subtracting the row left behind. The algorithm is denoted SWAHE (Sliding Window Adaptive Histogram Equalization) by the original authors. The computational complexity of histogram calculation is then reduced from O(N²) to O(N) (with N = pixel width of the surrounding rectangle); and since there is no tiling a final interpolation step is not required. Interpolation artifacts (a) shows the original input image with a linear gray value ramp. (b) shows the result of Interpolated AHE, with banding artifacts due to the interpolation. (c) shows the smooth result of exact CLAHE. The artifacts on the left and right are due to border treatment and thus inevitable Härtinger, P., Steger, C. Adaptive histogram equalization in constant time. J Real-Time Image Proc 21, 93 (2024). https://doi.org/10.1007/s11554-024-01465-1 References ⚫ Wilhelm Burger and Mark J. Burge, Digital Image Processing, Springer, 2008 ⚫ University of Utah, CS 4640: Image Processing Basics, Spring 2012 ⚫ Gonzales and Woods, Digital Image Processing (3rd edition), Prentice Hall ⚫ Digital Image Processing slides by Brian Mac Namee ⚫ CS 545/ECE 545 Digital Image Processing, Spring Semester 2014, Worcester. Polytechnic Institute ⚫ CS 589-04 Digital Image Processing, New Mexico Tech  Süleyman S. Narlı | DL Architecture for COVID-19 Diagnosis Impact of Local Histogram Equalization on Deep Learning Architectures for Diagnosis of COVID-19 on Chest X-rays Süleyman Serhan Narlı1, Gokhan Altan1 1 Iskenderun Technical University, Hatay, Turkey Corresponding author: Süleyman Serhan Narlı (e-mail: [email protected]). Article Info ABSTRACT Deep Learning (DL) is one of the most popular Machine Learning (ML) algorithms with feature learning capabilities. Its use is becoming widespread Received: January 04, 2021 day by day due to its high-performance in classification in various fields, including Accepted: April 03, 2021 Published: April 26, 2021 medical image processing. DL is inspired by an advanced neural network structure and includes many parameters. In consequence of its high performance, it is used in the classification of many diseases. DL algorithms, which are frequently used in image processing, classify the pixels on the images by convolutional progress in different layers. Before learning the significant pixels in supervised ways, it can be Keywords ensured that the classification is more successful with different pre-processing Deep Learning methods. In this study, the effect of DL architectures on COVID19 was investigated Convolutional Neural Networks Local Histogram Equalization using Local Histogram Equalization (LHE). Chest x-ray images were examined with- Medical Image Analysis and without-LHE to determine the impact of disk factor on transfer learning. The Chest X-Ray dataset consists of COVID-19, Pneumonia, and normal chest x-ray images. Chest x- rays were segmented into two parts of right lung lobe and the left lung lobe. The effect on the classification performance of transfer learning was observed by applying different disk value for LHE. The experiments were evaluated on the different pre- trained DL architectures, such as VGG16, AlexNet, and Inception model. I. INTRODUCTION separate feature extraction stage. In conventional image Image processing is preparing the image for classification processing techniques, Convolutional Neural Networks and segmentation by applying advanced mathematical (CNN) has started to be used more than the handcrafted solutions and signal processing techniques. A two- techniques, typical classifier deep neural networks, such as dimensional matrix generally considers the image with three AlexNet (Krizhevsky et al., 2012), VGGNet (Simonyan and color channels (RGB-Red Green Blue), and each pixel in Zisserman, 2015), or GoogLeNet (Szegedy et al., 2015) are images is presented by a mathematical value between 0- 255. fed directly by images. The sequence in DL generates class Many algorithms for classification in image processing rely on probability. Whereas extracting the probabilistic prediction, the detection and extraction of local features in images. the experiments are performed by using distinct datasets for Conventional image processing tries to obtain significance by training and validation. Generally, a CNN structure is formed characterizing the meaningful features in the image. A as; The raw image is fed as input to the first layer of the handcrafted method has been developed to solve the problem network by allowing each layer to calculate its respective of feature extraction for classification. Afterward, the function on the output of the preceded layer. One of the most elaborated handcrafted techniques were used in classification studied classifications is handwritten number recognition, in and segmentation. However, applying pattern recognition which a neural network takes the image of a single systems with handcrafted techniques for large datasets is a handwritten digit. It must decide which of the ten digits are challenging process. There is no need to use handcrafted displayed on the images (Li et al., 2014). Deep learning has a techniques in deep learning (Lowe, 2004); It uses many widespread-use in a wide variety of machine learning tasks, (deeper) unit layers with highly optimized algorithms, including image classification, speech recognition, medical architectures, and regularizations. Feature learning and image processing, natural language processing, and more. transfer learning, which are novel techniques in Deep There are many novel types of deep learning research Learning (DL), are used in pattern recognition without a (Goodfellow et al., 2016; LeCun et al., 2015; Pouyanfar et al., VOLUME 02, NO: 01. 2021 11  Süleyman S. Narlı | DL Architecture for COVID-19 Diagnosis. 2018; Wu et al., 2019; Zhang et al., 2018; Zhou et al., 2018; architecture has been presented in (Rasheed et al., 2021a) to Zhu et al., 2017). These studies examined deep learning distinguish COVID-19 patient among normal by analyzing techniques from different perspectives, with the applications chest X-rays. Many studies were performed on DL-based of medical image analysis (Litjens et al., 2017), natural medical image applications. Examples include SegNet language processing (Young et al., 2018), speech recognition (Badrinarayanan et al., 2017), DenseU-Net (Li et al., 2018), systems (Zhang et al., 2018), and remote sensing (Zhu et al., and CardiacNet (Mortazi et al., 2017) (Baltruschat et al., 2019) 2017). Deep Neural Networks (DNN) consists of several in detecting Pneumonia in CXRs. A more detailed review layers of nodes; different architectures have been developed to about COVID-19 detection can be found in (Rasheed et al., solve various issues in different fields of use. Whereas CNN 2020) and (Rasheed et al., 2021b). In this study, pre- is often used for computer vision and image recognition, processing was performed using the Local Histogram recurrent neural network (Cho et al., 2014) is often used for Equalization (LHE) method to make the tissues and masses in time-series. The most common DNN architectures are this; 1) the lung images more visible before feeding the CNN for CNN, 2) Deep Autoencoders, 3) Long Short Term Memory pathology detection on chest x-rays. Afterward, the (LSTM), 4) Restricted Boltzmann Machine (RBM). All these performance of CNN and the effect of LHE with CNN on the Deep architectures use backpropagation in the training stage detection of COVID-19, Pneumonia, and normal images were of supervised learning. Backpropagation uses gradient- evaluated. After LHE, it is aimed to increase the performance descent for error minimization, adjusting the weights by combining similar tissues with CNN in the training stage. according to the partial derivative of the error function for each Lung images were cropped and separated as right and left lobe weight. Model training can be broadly divided into three types, images, and only lung sections were inhaled. Thus, nonlung including supervised, unsupervised, and semi-supervised. sections were eliminated from chest x-rays and enabled the Supervised learning uses labeled data to train the network, CNN architecture to be trained only on the lung sections for whereas unsupervised learning does not need a labeled dataset. experimented architectures. The remaining of the paper is Accordingly, there is no feedback-based learning. In organized as follows. The CNN architecture, chest x-rays unsupervised learning, neural networks are pre-trained using database, transfer learning on AlexNet, and LHE technique are manufacturing models such as RBMs and then fine-tuned detailed in Materials and Methods. The experimental setup using standard supervised learning algorithms on pre-trained and the achievements for various CNN architectures with- and weights. Afterward, it is tested by a separate dataset to identify without-LHE were presented and evaluated in experimental patterns or classification results. transfer learning is adapting results. The advantages of the proposed technique, state-of- the weights of trained models for a different dataset. In DL, the-art, comparison considering various parameters with pruning is mainly used to developing a smaller and more literature are evaluated and discussed in the Discussion efficient neural network model. This technique aims to section. optimize the model by eliminating the values of weight tensors and obtaining a less time-consuming, computationally cost- II. MATERIALS AND METHODS effective model in training. Many studies have been performed A. CONVOLUTIONAL NEURAL NETWORKS in the field of medical image processing, such as CT images. CNNs are a deep network model consisting of sequential 3-dimensional Deep architectures were used to automatically convolutional layers and fully connected layers. The segment abdominal CT to identify arteries, portal veins, liver, convolutional layers serve as feature extractors and learn the spleen, stomach, gallbladder, and pancreas in each multi- various representations of the input images. The output of organ image (Roth et al., 2018). Deep convolution and deep convolutional layers is arranged in feature maps. Each neuron belief networks, a type of RBM network, conditional random in a feature map has a receptive field that connects to a layer field, and structured support vector machine were utilized to of neurons in the previous layer via a trainable set of weights separate breast masses from mammograms on mammograms (LeCun et al., 2015). Inputs are combined with learned (Navab et al., 2015). Various methods were presented in the weights to create a new feature map, and folded results are literature to detect Pneumonia using chest x-ray images. Some transferred via rectified linear units (ReLU) (Nair et al., 2010), of these methods use a machine learning algorithm as the which is a nonlinear activation function. The weights of each classification technique and deep learning techniques for neuron in a feature map are equally constrained; however, feature extraction and classification (García et al., 2020). different feature maps within the same convolution layer have Similarly, (Antin et al., 2017) used logistic regression as a different weights so that various features can be extracted at basic model for pneumonia detection using x-rays using a 121- each location. The convolution layer consists of using various layer dense convolutional network (DenseNet). In another dimensions and numbers of filters to represent different study, Rajpurkar et al. developed a model that diagnoses properties of the inputs. The depth of the CNN is defined by Pneumonia with a high accuracy rate; ChexNet developed a the number of filters in each transform layer (Ciresan et al., 121-layer CNN model that analyzes chest x-ray image and 2011). Deep learning models also include the pooling layer classified it bilaterally, and determined the probability of apart from the convolution layer; the purpose of this layer is to Pneumonia (Rajpurkar et al., 2017). A deep-learning based reduce the spatial resolution of the feature maps and thus 12 VOLUME 02, NO: 01. 2021 Süleyman S. Narlı | DL Architecture for COVID-19 Diagnosis. achieve spatial invariance to enter distortions and translations. layer and fifth layer, then three fully connected layers, It is used to propagate the average of all values of a small respectively (Krizhevsky et al., 2012). In this study, transfer region of an image to the next layer (LeCun et al., 2015), learning was applied using the AlexNet architecture, and the (Ranzato et al., 2007). After various sequential convolution chest x-ray images were classified into multi classes. Transfer layers and pool layers, the feature maps are stacked on top of learning is a method of retraining a pre-trained model on a new each other, and the fully connected layers are connected to dataset. The model is applied to the new dataset by fine-tuning flattened feature maps. It interprets these layer representations the model in transfer learning. In this study, we adapted and performs the classification function (Hinton et al., 2012), transfer learning into the chest x-ray database using the (Zeiler et al., 2014). It is standard to use the softmax function AlexNet architecture. The number of filters in the convolution as an output layer of fully connected layers (Krizhevsky et al., layers of the model was reduced by pruning. Thus 2012). performance improvement was reported on pruned AlexNet architecture. Since AlexNet architecture is a model created for B. CHEST X-RAY DATABASE a large dataset, it is well-adapted to the small-scale dataset. Chest x-ray images are a dataset obtained from real patients The AlexNet architecture was pruned, and the supervised by chest x-ray examination and available in the clinical PACS stage was trained using a single fully connected layer instead database at the National Institutes of Health Clinical Center. of the last three fully connected layers. Thus, it was ensured The chest x-ray dataset includes 112,120 frontal view medical that there was not much cohesion in the training of the model. images of 30,805 unique patients with fourteen pulmonary The proposed architecture is depicted in Figure (2). diseases (each image can have multiple labels). Fourteen common thoracic pathologies are atelectasis, consolidation, D. LOCAL HISTOGRAM EQUALIZATION infiltration, pneumothorax, edema, emphysema, fibrosis, Automatic contrast enhancement is one of the common effusion, Pneumonia, pleural thickening, cardiomegaly, operations performed on visual data to reveal confidential nodule, mass, and hernia (Xiaosong et al., 2017). This dataset detail. A histogram is a graph showing pixel density. Global is publicly available on the Kaggle platform1. Similar datasets: histogram equalization (GHE) is generally a global tone Montgomery County X-ray Set, Shenzhen Hospital X-ray Set mapping process, allowing the grey level of each pixel to be (Jaeger, 2014), ChexNet (Rajpurkar et al., 2017). Pneumonia- recreated by calculating the overall histogram of the image. diagnosed lung images and healthy lung images were selected However, these processes fail to increase contrast in both dark from the NIH Chest x-ray dataset. The chest x-rays with and bright image regions at the same time. Especially, small COVID-19 were chosen from the COVIDx dataset. This bright spots are hardly visible after such a global operation. dataset is the largest open-access dataset in terms of the number of COVID-19 positive patient cases2. Many datasets are also publicly available in (Cohen et al., 2020). The chest x-ray can indicate pathological sections, such as edema, nodule, and infiltration in the lungs according to different intensity colors, making diagnosis difficult in the early stages due to the noisy image. In such cases, thanks to the filters applied to the image and localizing the lung regions by segmentation on the images, and the noise can be reduced by selecting certain regions. The performance of the dataset trained with deep learning can be increased. When we examine the publications in this area, the medical images can be processed with machine learning methods. The diagnosis of pathology for many lung diseases can be performed with FIGURE 1. Different LHE (disk) values the sensitivity of the doctor. Automatized techniques can also Local histogram equalization (LHE) was performed in a play an essential role in identifying many pathologies and floating window with this adjustable window size (radius or early diagnosis with the enhancements of DL algorithms and disk) to solve this problem. Histogram equalization is applied architectures. independently to small areas of the image, thus preserving the contrast adjustment for different regions of the image (Jaehne, C. ALEXNET AND TRANSFER LEARNING 1991; Caselles et al., 1999; Y. Kim et al., 2001). In this study, AlexNet is a DL architecture created by Alex Krizhevsky et chest x-ray images were pre-processed using LHE. Figure (1) al. (Krizhevsky et al., 2012) for ImageNet large visual indicates the lung images for different LHE window sizes recognition in 2012. It is an essential, simple, and effective (disk) values. CNN architecture consisting mainly of gradual stages such as III. EXPERIMENTAL RESULTS convolution layers, pool layers, corrected linear unit (ReLU). AlexNet architecture was proposed by Alex Krizhevsky et Specifically, AlexNet has five convolutional layers: first layer, al. in 2012. The total data size in this study consists of 2856 second layer, third layer, and fourth layer, followed by pool images for each model training, which is a small-scale dataset VOLUME 02, NO: 01. 2021 13  Süleyman S. Narlı | DL Architecture for COVID-19 Diagnosis. for the AlexNet architecture. Therefore, while the data is Tensorflow framework with python programming language adapted to the AlexNet architecture with transfer learning, the was chosen to train the model (GPU: Nvidia RTX 2060). number of filters in the convolution layers has been changed. We calculated accuracy, precision, recall, and f1 Score were The model has been optimized according to the training used to evaluate the performance of the model. The accuracy performance. By cropping the chest x-ray images, two lung of a test is the ability to distinguish between patient and lobes were cropped from each chest x-ray separately. Thus, healthy cases accurately. Recall, the sensitivity of a test is its 1000 healthy lung images, 1000 lung images with Pneumonia, ability to identify patient cases accurately. It is the metric of and 856 chest x-rays with COVID-19 were utilized in the the true positive rate in diseased cases. Precision is the metric to test the models’ ability to accurately identify healthy cases TABLE I CONFUSION MATRIX RESULTS FOR DIFFERENT LHE DISK and calculate the true negative rate in healthy cases. Confusion VALUES Matrix results for different LHE Disk values are presented in Confusion Table I. Normal Pneumonia Covid-19 Matrix Normal 170 3 7 𝑇𝑃 + 𝑇𝑁 Without 𝑂𝑣𝑒𝑟𝑎𝑙𝑙 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (1) LHE Pneumonia 4 200 1 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 Covid-19 25 10 150 Normal 150 8 24 LHE (disk = 20) Pneumonia 0 190 20 𝑇𝑃 Covid-19 18 14 150 𝑅𝑒𝑐𝑎𝑙𝑙 = (2) 𝑇𝑃 + 𝐹𝑁 Normal 160 6 13 LHE Pneumonia 0 200 5 (disk = 40) Covid-19 15 4 160 𝑇𝑃 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (3) TABLE II 𝑇𝑃 + 𝐹𝑃 COMPARISON OF RELEVANT DL-BASED DISEASE CLASSIFICATION STUDIES IN TERMS OF DIFFERENT DISK 2𝑇𝑃 VALUES FOR LHE 𝐹1𝑠𝑐𝑜𝑟𝑒 = (4) Disk Data Accuracy Precision Recall F1- 2𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁 Value (Train/Test) Score True positive (TP), number of cases correctly identified Without 2284/572 91% 91% 91% 91% patients; false positive (FP), number of cases incorrectly LHE identified as patients; true negative (TN), the number of cases 20 2284/572 85% 85% 85% 85% correctly identified as healthy; and the number of false 40 2284/572 92% 92% 92% 92% negative (FN) cases incorrectly identified as healthy. analysis. Two floating window parameters (disk) were used IV. DISCUSSION for LHE with 20 and 40. The model has been optimized by COVID-19 analysis of many researching areas has gained fine-tuning the parameters on the AlexNet model. Instead of importance to ease the diagnosis and detection rate of disease the last three fully connected layers in the AlexNet for a pandemic. Consequently, advancing image processing architecture, one fully connected layer is used. Thus we techniques and proposing robust CADx systems are the main prevented overfitting and reducing the number of the focus of researchers. The manuscripts for assessments of classification parameters by pruning. As the AlexNet COVID-19 and using automatized methods for abnormality architecture, 5 convolution layers were used in the pruned detection within various diagnostic tools indicate the model, and the number of filters for each layer are determined popularity of the field in 2020. Whereas high classification in this way for 32, 64, 128, 256, and 512, respectively. The performances for diagnosis of COVID-19 were reported using dataset is split into three folds with 20%, 20%, and 60% for complex Deep learning architectures, applying simple pre- model testing, validation, and training, respectively. The batch processing stages to the evaluation process reached high size is 32, and the epoch number is 50 in the training of the enough achievements using pruned and simple Deep fully connected layers. Early stopping was used to prevent architectures. The main ideas are using the generated overfitting during supervised learning. In the last layer, three representations instead of many convolutional layers, multi-classes are used for classification using the softmax transmitting feature learning within resembled instances with function (COVID-19, Pneumonia, and Normal). We set LHE, and adapting the capability of feature transferring 0.85e*5 for the learning rate in the Adam optimizer. In this between adjacent layers to identify chest x-rays with healthy, manner, the hyper-parameters in the model were arranged. COVID-19, and pneumonia pathology. The total number of classification parameters was 2,072,899. 14 VOLUME 02, NO: 01. 2021 Süleyman S. Narlı | DL Architecture for COVID-19 Diagnosis TABLE III COMPARISON OF RELEVANT DL-BASED COVID-19 DISEASE DETECTION STUDIES IN TERMS OF ALGORITHMS AND PERFORMANCES Related Methods Classifier Covid-19 Test ACC F1 Score F1-score for Works Sample (%) (%) COVID-19 (%) Wang et al. COVID-NetCNN CNN 100 93.3 93.13 94.78 Karthik et al. Customized CNN with the CNN 112 97.94 96.90 97.20 distinctive filter learning module Ozturk et al. DarkNet-19 CNN 25 87.02 88.0 88 Khan et al. CoroNet net CNN 70 89.6 89.8 95.61 Apostolopoulos et Transfer learning with MobileNetV2 CNN 222 94.72 93.80 90.50 al. Farooq et al. Transfer learning with ResNet50 CNN 68 96.23 96.88 100 Hemdan et al. COVIDX-Net CNN 25 90 - - Narin et al. ResNet-50 CNN 100 98 - - Our Study AlexNet Model without LHE CNN 572 91 91 87 Our Study AlexNet Model With LHE Disk 20 CNN 572 85 85 80 Our Study AlexNet Model With LHE Disk 40 CNN 572 92 92 90 Table III presents the comparison of our study with the deep Karthik et al. proposed a different deep learning model to learn learning algorithms developed to diagnose COVID-19. Wang specific filters within a single convolutional layer to identify et al. reported an accuracy rate of 93.3% for their proposed Pneumonia classes. Their proposal achieved an F1 score rate COVID-Net model using cropping, translation, rotation, of 97.20% for 112 chest x-rays with COVID-19. horizontal rotation, zoom on chest x-rays (Wang et al., 2020). Moreover, the lung regions were divided into sections by ResNet50, ResNet101, ResNet152, InceptionV3 and applying a pre-trained algorithm, and the segmentation Inception-ResNetV2. He applied three different binary process was applied (Karthik et al., 2020). Ozturk et al. classifications with four classes (COVID-19, normal presented the DarkNet model, a novel model for detecting (healthy), viral Pneumonia, and bacterial Pneumonia) using COVID-19, which consists of 17 convolutional layers for 5-fold cross-validation. Among the other four models used, binary classification (COVID-19 and no finding), and multi- the pre-trained ResNet50 model had the highest class classification (no finding, COVID-19, and classification performance for different datasets with an Pneumonia). The model achieved classification accuracy accuracy rate of 99.5% (Narin et al., 2020). The state-of-the- rates of 98.08% and 87.02% for binary classification and art indicates that the pre-trained CNN architectures, pruned multi-class classification (Ozturk et al., 2020). Khan et al. net-works, and using detailed feature learning stages can reported the CoroNet model based on the pre-trained separate COVID-19from healthy and Pneumonia. Using Xception CNN architecture and achieved a classification hybrid methods in resembling stage of deep architectures, accuracy of 95% for three classes (COVID-19, Pneumonia, pruning the complex pre-trained architectures with transfer and normal) (Khan et al., 2020). Apostolopoulos et al. learning, and proposing lightweight architectures with reached an accuracy rate of 96.78% using transfer learning simplistic image processing approaches can advance the on MobileNetV2 (Apostolopoulos et al., 2020). Farooq et al. deep models. The proposed pruned CNN architecture with developed the COVID-ResNet model by fine-tuning the LHE is a powerful trajectory to get high enough results with ResNet50 model with transfer learning and achieved an easy-adaptable real-time applications over to the state-of-art. accuracy rate of 96.23% in the COVIDx dataset The proposed model possesses the capabilities of pre-trained (Muhammad et al., 2020). Hemdan et al. proposed a novel detailed CNN architectures by feeding the input with model COVIDX-Net based on seven different architectures resembled chest x-rays before the layer-wise feature of DCNNs; VGG19, DenseNet201, InceptionV3, learning. LHE for generating smoothed representation ResNetV2, InceptionResNetV2, Xception, and provided high enough classification performances using MobileNetV2. They reported a performance rate range for pruned AlexNet architecture. Therefore, simplifying the various pre-trained architectures between 80-90% (Hemdan deep architectures reduces the training time for COVID-19 et al., 2020). Narin et al. evaluated five pre-trained identification and integrates the shallow architectures for convolutional neural network-based models including. even real-time applications in embedded systems. VOLUME 02, NO: 01. 2021 15 Süleyman S. Narlı | DL Architecture for COVID-19 Diagnosis V. CONCLUSION MICCAI 2015. MICCAI 2015. Lecture notes in computer science, vol. 9349 (2015), pp. 605-612 The study contains the novelty of applying LHE for Digital Image Processing-Concepts, Algorithms, and Scientific generating resembling representations for increasing the Applications, Jaehne. B, Springer, Berlin 1991 generalization capability of CNN with shallow Farooq Muhammad, Hafeez Abdul. 2020. Covid-resnet: a deep learning architectures. Whereas many studies focus on just framework for screening of covid19 from radiograph. arXiv preprint arXiv:2003.14395. classification performances in supervised training and Goodfellow, I., Bengio, Y., Courville, A. (2016). Deep learning. modeling feature learning stages, the proposal is a pioneer Cambridge: MIT press. study with LHE, which has a common use on medical H. Li, R. Zhao, and X. Wang. Highly efficient forward and backward propagation of convolutional neural networks for pixelwise images. The necessity of big data for the training of CNN classification. 2014, arXiv:1412.4526v2 [cs.CV]. architectures is overcome using LHE as a resembling H.R. Roth, et al. Deep learning and its application to medical image procedure using a middle-scale chest x-ray dataset. The segmentation (2018), pp. 1-6 Hemdan E.E.-D., Shouman M.A., Karar M.E. COVIDX-Net: A proposed model has reached high identification rates for framework of deep learning classifiers to diagnose covid-19 in X- multi-class diseases using is a basic, pruned architecture. It ray images (2020) arXiv preprint arXiv:2003.11055 is easy-to-integrate architecture for various types of medical Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I., Salakhutdinov, R. R. (2012). Improving neural networks by image analysis. Using novel visualization techniques and preventing co-adaptation of feature detectors. arXiv 1207.0580. localization algorithms on chest x-rays has a possible I. M. Baltruschat, H. Nickisch, M. Grass, T. Knopp, and A. Saalbach, contribution to Deep Learning algorithms. Although CNN “Comparison of Deep Learning Approaches for Multi-Label Chest architectures don’t need additional feature extraction and X-Ray Classification,” Sci. Rep., vol. 9, no. 1, pp. 1–10, Dec. 2019, doi: 10.1038/s41598-019-42294-8. pre-processing procedures, it has limitations in detecting J. Y. Kim, L. S. Kim, and S. H. Hwang, “An advanced contrast small pathologies with low generalization capabilities. In the enhancement using partially overlapped sub-block histogram view that the classification assignments for modeling equalization,” IEEE Trans circuits systems and video technology, vol. 11, no. 4, pp. 475-484, Apr. 2001. architectures in CNN limited the variance in the COVID-19 Jaeger S, Candemir S, Antani S, Wáng YX, Lu PX, Thoma G. Two public analysis on chest x-rays, feeding the architectures with chest X-ray datasets for computer-aided screening of pulmonary generative representations is a novel approach for diseases. Quant Imaging Med Surg. 2014 Dec;4(6):475-7. doi: 10.3978/j.issn.2223-4292.2014.11.20. PMID: 25525580; PMCID: minimizing the loss and increasing performance with even PMC4256233. shallow models. In further analysis, evaluating various K. Cho. (2014). ‘‘Learning phrase representations using RNN localization parameters and experimenting with various encoderdecoder for statistical machine translation.’’ [Online]. lightweight architectures will enhance the performance Available: https://arxiv.org/abs/1406.1078 K. Simonyan and A. Zisserman. Very deep convolutional networks for capabilities on both left, right, and both lungs, separately. large-scale image recognition. 2015, arXiv:1409.1556v6 [cs.CV]. Karthik R, Menaka R, M H. Learning distinctive filters for COVID-19 REFERENCES detection from chest X-ray using shuffled residual CNN. Appl Soft 1 https://www.kaggle.com/nih-chest-xrays/data Comput. 2021;99:106744. doi:10.1016/j.asoc.2020.106744 2 www.kaggle.com/tawsifurrahman/covid19-radiography-database Khan A.I., Shah J.L., Bhat M.M. Coronet: A deep neural network for A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification detection and diagnosis of COVID-19 from chest x-ray images. with deep convolutional neural networks. In NIPS, 2012. Comput. Methods Programs Biomed. 2020 A. Mortazi, R. Karim, K. Rhode, J. Burt, and U. Bagci, “CardiacNET: LeCun, Y., Bengio, Y., Hinton, G. (2015). Deep learning. Nature, 521, Segmentation of left atrium and proximal pulmonary veins from 436–444. MRI using multi-view CNN,” in Lecture Notes in Computer LeCun, Y., Bengio, Y., Hinton, G. (2015). Deep learning. Nature, Science (including subseries Lecture Notes in Artificial Intelligence 521(7553), 436–444. and Lecture Notes in Bioinformatics), 2017, vol. 10434 LNCS, pp. Litjens, G., Kooi, T., Bejnordi, B., Setio, A., Ciompi, F., Ghafoorian, M., 377–385, doi: 10.1007/978-3-319-66185-8-43. et al. (2017). A survey on deep learning in medical image analysis. Antin, B.; Kravitz, J.; Martayan, E. Detecting Pneumonia in Chest X-rays Medical Image Analysis, 42, 60–88. with Supervised Learning; Semanticscholar Org.: Allen Institute for Lowe, D.G. Distinctive Image Features from Scale-Invariant Keypoints. Artificial intelligence, Seattle, WA, USA, 2017. International Journal of Computer Vision 60, 91–110 (2004). Apostolopoulos I.D., Mpesiana T.A. Covid-19: automatic detection from https://doi.org/10.1023/B:VISI.0000029664.99615.94 X-ray images utilizing transfer learning with convolutional neural Luján-García, J.E.; Yáñez-Márquez, C.; Villuendas-Rey, Y.; Camacho- networks. Phys. Eng. Sci. Med. 2020 Nieto, O. A Transfer Learning Method for Pneumonia C. Szegedy, W. Liu, Y. Jia, et al. Going deeper with convolutions. In 2015 Classification and Visualization. Appl. Sci. 2020, 10, 2908. IEEE Conference on Computer Vision and Pattern Recognition, Nair, V., Hinton, G. E. (2010). Rectified linear units improve restricted pages 1–9, 2015. Boltzmann machines. In Proceedings of the 27th International Cireşan DC, Meier U, Masci J, Gambardella LM, Schmidhuber J,(2011) Conference on Machine Learning (pp. 807–814). N.p.: International Flexible, high performance convolutional neural networks for Machine Learning Society. image classification. In: IJCAI International Joint Conference on Narin A., Kaya C., Pamuk Z. Automatic detection of coronavirus disease Artificial Intelligence. https://doi.org/10.5591/978-1-57735-516- (COVID-19) using X-ray images and deep convolutional neural 8/IJCAI11-210. networks (2020) Cohen, J. P., Morrison, P.,Dao, L. COVID-19 image data collection. Ozturk T., Talo M., Yildirim E.A., Baloglu U.B., Yildirim O., Rajendra arXiv:2003.11597 (2020). Acharya U. Automated detection of COVID-19 cases using deep Coronavirus World nHealth Organization: neural networks with X-ray images. Comput. Biol. Med. 2020;121 https://www.who.int/healthtopics/coronavirus,2020. Pouyanfar, S., Sadiq, S., Yan, Y., Tian, H., Tao, Y., Reyes, M. P., et al. Deep learning and structured prediction for the segmentation of mass in (2018). A survey on deep learning: Algorithms, techniques, and mammograms, N. Navab, J. Hornegger, W. Wells, A. Frangi (Eds.), applications. ACM Computing Surveys, 51(5), 92:1–92:36. Medical image computing and computer-assisted intervention – Rajpurkar, P.; Irvin, J.; Zhu, K.; Yang, B.; Mehta, H.; Duan, T.; Ding, D.; Bagul, A.; Langlotz, C.; Shpanskaya, K.; et al. Chexnet: VOLUME 02, NO: 01. 2021 16 Süleyman S. Narlı | DL Architecture for COVID-19 Diagnosis. Radiologist-level pneumonia detection on chest x-rays with deep X-ray images. Sci Rep 10, 19549 (2020). learning. arXiv 2017, arXiv:1711.05225. https://doi.org/10.1038/s41598-020-76550-z Ranzato, M. A., Huang, F. J., Boureau, Y., LeCun, Y. (2007). Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., Yu, P. S. (2019). A Unsupervised learning of invariant feature hierarchies with comprehensive survey on graph neural networks. applications to object recognition. In Proceedings IEEE Conference arXiv:1901.00596. on Computer Vision and Pattern Recognition (pp. 1–8). Los X. Li, H. Chen, X. Qi, Q. Dou, C. W. Fu, and P. A. Heng, “HDenseUNet: Alamitos, CA: IEEE Computer Society. Hybrid Densely Connected UNet for Liver and Tumor Rasheed, J., Jamil, A., Hameed, A.A., Aftab, U., Aftab, J., Shah, S.S., Segmentation from CT Volumes,” IEEE Trans. Med. Imaging, vol. Draheim, D. A survey on artificial intelligence approaches in 37, no. 12, pp. 2663–2674, Dec. 2018, doi: supporting frontline workers and decision makers for the COVID- 10.1109/TMI.2018.2845918. 19 pandemic. Chaos, Solitons & Fractals 141, 110337, (2020). Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, https://doi.org/10.1016/j.chaos.2020.110337 MohammadhadiBagheri, Ronald Summers, ChestX-ray8: Hospital- Rasheed, J., Hameed, A.A., Djeddi, C., Jamil, A., Al-Turjman, F. A scale Chest X-ray Database and Benchmarks on Weakly- machine learning-based framework for diagnosis of COVID-19 Supervised Classification and Localization of Common from chest X-ray images. Interdiscip Sci Comput Life Sci 13, 103– ThoraxDiseases, IEEE CVPR, pp. 3462-3471,2017 117 (2021a). https://doi.org/10.1007/s12539-020-00403-6 Young, T., Hazarika, D., Poria, S., Cambria, E. (2018). Recent trends in Rasheed, J., Jamil, A., Hameed, A.A., Al-Turjman, F., Rasheed, A. deep learning based natural language processing. IEEE COVID-19 in the Age of Artificial Intelligence: A Comprehensive Computational Intelligence Magazine, 13(3), 55–75. Review. Interdiscip Sci Comput Life Sci (2021b). Zeiler, M. D., Fergus, R. (2014). Visualizing and understanding https://doi.org/10.1007/s12539-021-00431-w convolutional networks. In Proceedings of the European Roosa K, Lee Y, Luo R, Kirpich A, Rothenberg R, Hyman JM, and et al. Conference on Computer Vision (pp. 818–833). Berlin: Springer. Real-time forecasts of the COVID-19 epidemic in China from Zhang, Z., Geiger, J., Pohjalainen, J., Mousa, A. E., Jin, W., Schuller, B. February 5th to February 24th, 2020. Infectious Disease Modelling, (2018d). Deep learning for environmentally robust speech 5:256-263, 2020. recognition: An overview of recent developments. ACM V. Badrinarayanan, A. Kendall, and R. Cipolla, “SegNet: A Deep Transactions on Intelligent Systems and Technology, 9(5), 49:1– Convolutional Encoder-Decoder Architecture for Image 49:28. Segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. Zhou, J., Cui, G., Zhang, Z., Yang, C., Liu, Z., Sun, M. (2018a). Graph 12, pp. 2481–2495, Dec. 2017, doi: 10.1109/TPAMI.2016.2644615 neural networks: A review of methods and applications. V. Caselles, J. L. Lisani, J. M. Morel, and G. Sapiro, “Shape preserving arXiv:1812.08434. local histogram modification,” IEEE Trans image proc,vol. 8, no. 2, Zhu, X., Tuia, D., Mou, L., Xia, G., Zhang, L., Xu, F., et al. (2017). Deep pp. 220-230, Feb. 1999. learning in remote sensing: A comprehensive review and list of Wang, L., Lin, Z.Q. Wong, A. COVID-Net: a tailored deep convolutional resources. IEEE Geoscience and Remote Sensing Magazine, 5(4), neural network design for detection of COVID-19 cases from chest 8–36. VOLUME 02, NO: 01. 2021 17  Available online at www.sciencedirect.com Available Available online online at at www.sciencedirect.com www.sciencedirect.com Available online at www.sciencedirect.com Available online at www.sciencedirect.com Procedia Computer Science 00 (2023) 000–000 ScienceDirect Procedia www.elsevier.com/locate/procedia Procedia Computer Computer Science Science 00 00 (2023) (2023) 000–000 000–000 Procedia Procedia Computer Computer Science Science 00 (2023) 225 (2023) 000–000 3021–3030 www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia 27th International Conference on Knowledge-Based and Intelligent Information & Engineering 27th International Systems (KES 2023)Intelligent Information & Engineering 27th International Conference Conference on on Knowledge-Based Knowledge-Based and and Intelligent Information & Engineering 27th International Conference on Knowledge-Based Systems (KES and Intelligent Information & Engineering 2023) Modified Histogram Equalization Systems (KES 2023) Systems (KES for 2023)Improved CNN Medical Modified Histogram Image Equalization for Improved CNN Medical Segmentation Modified Histogram Equalization for Improved CNN Medical Image Segmentation Image Shoffan Segmentation Saifullah a,b,∗ , Rafał Dreżewskia a,b,∗ aa a Institute Shoffan Shoffan Saifullah ,, Rafał Dreżewski a,b,∗ of Science and Technology, Kraków, Shoffan Saifullah , Rafał Rafał Dreżewski of Computer Science, AGH University Poland a,b,∗ a Saifullah b Department of Informatics, Universitas Pembangunan Dreżewski Nasional Veteran Yogyakarta, Yogyakarta, Indonesia a Institute of Computer Science, AGH University of Science and Technology, Kraków, Poland a Institute of Computer Science, AGH University of Science and Technology, Kraków, Poland a Institute of Computer Science, AGH University of Science and Technology, Kraków, Poland b Department b Department of of Informatics, Informatics, Universitas Universitas Pembangunan Pembangunan Nasional Nasional Veteran Veteran Yogyakarta, Yogyakarta, Yogyakarta, Yogyakarta, Indonesia Indonesia b Department of Informatics, Universitas Pembangunan Nasional Veteran Yogyakarta, Yogyakarta, Indonesia Abstract Abstract This research aims to improve the performance of convolutional neural network (CNN) in medical image segmentation that will Abstract Abstract detect specific parts of the body’s anatomical structures. Medical images have drawbacks, such as the image’s variability, quality, This This research research aims aims to improve improve the performance of of convolutional neural network network (CNN) in in medical image segmentation that that will and This complexity. research Weto aims toofdeveloped improve the the performance image preprocessing performance convolutional of convolutional neuralHistogram scenarios using neural network (CNN) Equalization (CNN) medical insuch (HE),image medical segmentation Contrast image Limited Adaptive segmentation that will will detect detect specific specific parts parts of the the body’s body’s anatomical anatomical structures. structures. Medical Medical images images have have drawbacks, drawbacks, such as as the the image’s image’s variability, variability, quality, quality, Histogram detect Equalization specific parts (CLAHE), ofdeveloped the body’simage and the anatomical hybrid approaches structures. (HE-CLAHE Medicalusing images and CLAHE-HE). have drawbacks, such (HE), We propose as the Contrast CNN with image’s variability, image quality, and and complexity. complexity. We preprocessing scenarios Histogram Equalization Limited Adaptive enhancement and complexity. for We image We developed developed imageand segmentation image preprocessing preprocessing scenarios using evaluate its performance scenarios using Histogram on Lung CT-Scan Histogram Equalization and Chest (HE), Equalization X-ray Contrast (HE), Limitedtotaled datasets, which Contrast Limited Adaptive 267 Adaptive Histogram Histogram Equalization Equalization (CLAHE), (CLAHE), and the the hybrid and ground hybrid approaches approaches (HE-CLAHE (HE-CLAHE and and CLAHE-HE). CLAHE-HE). We We propose propose CNN CNN with with image image and 3616 Histogram images, respectively, Equalization and (CLAHE), and had truth. The experimental results indicate that the optimal cumulative distribution enhancement enhancement for image forvalue image segmentation andthe hybriditsapproaches evaluate its performance (HE-CLAHE on Lung and CLAHE-HE). CT-Scan and and Chest We datasets, X-ray propose CNN with datasets, which image totaled 267 function enhancement(CDF) for image of segmentation HE is 0 to 39,and segmentation andevaluate and the clip limit evaluate performance of CLAHE on its performance on LungCNN is 0.01. Lung CT-Scan produces CT-Scan Chest and that Chest X-ray the best segmentation X-ray datasets, which totaled with the which 267 addition totaled 267 and and 3616 3616 images, respectively, images, respectively, and had and method ground had ground truth. truth. The Thetheexperimental experimental results results indicate indicate that the optimal the optimal cumulative cumulative distribution distribution of the CLAHE-HE and 3616(CDF)images, approach. respectively, This and had can ground increase truth. The accuracy experimental by 1.23 results percentage indicate points that the (training) optimal and 3.22 cumulative percentage distributio