A New Model for Image Segmentation Based on Deep Learning

Egypt Abstract —Image segmentation is main point in computer vision (CV) and image processing (IP), that are used routinely in the fields of medicine and surgery training tools. Segmenting images and converting into a model that depends on work by the different algorithms from analysis DICOM files to convert to three-dimensional models. This paper describes a combination of two fields of solving segmentation problem to convert through the workflow of a hybrid algorithm structure Convolutional neural network (CNN, Active Contour & Deep Multi-Planar) based on seg3d2 to switch DICOM medical rays “Digital Imaging and Communications in Medicine” into a 3Dimintional model, using data from active contour to be the input of deep learning. the result of the pre-processing from DICOM raw images, each image contains edges and image size =256 X 256 pixel, which through adjustment and control we can create multiple results for output using Active Contour, by resizing the threshold frames and gray-scale image, and show liver 3D-model Deep architecture, it is through the CNN which the images of the three axes X, Y, and Z (three orthogonal) (coronal = X, sagittal = Y, axial = Z = 1) are determined and matched with a real image of the body, the area required to be determined, and edits the contrast using a histogram. This research will be using are human liver DICOM images and is divided into two stages (medical image segmentation - retinal model optimization), to help surgeons to study the patient’s condition with accuracy and efficiency through the use of mixed reality technology in liver surgery [living donor liver transplantation (LDLT)], all implement by Seg3D2 and


DICOM & Big data classifications
DICOM files contain a lot of information storage known as BIG DATA, most of which is not required. We need to define the working area (the range of interest within the image) to get useful information. DICOM is closely used by hospitals and surgeons such as Brain and liver surgery clinics. DICOM consists of multi-layers images that are combined through a specific system to show radiograph results.
Big data (BD) [1] is a huge collection of information (either structured, semi-structured, or unorganized) on archival units. Big information creates an incentive to build systems of the store and prepares data that cannot be split using traditional processes [2]. Big information do not containing sample data only the same numbers, dates, and strings. BD is Bulky information, contains a Geographic Information System (GIS) 3D model, medical image information, audio, video, DICOM documents, networking documents, and online history [3]. Despite the huge volume of information stored on the electronic cloud that can reach a petabyte or and Exabyte of information, it is related to the process of stability and stability of information, the extent of its exchange, and the implementation of important activities on time, and the extent to which it is possible to analyze and display information in "DICOM" files with huge parameters and stored inside several Layers [4][5] [6] .

Medical Img.Seg. (Image-segmentation)
Img.Seg. is a complex part of the pattern recognition system and the main part & 1st step in image analysis is one of the most complicated steps in image processing and determines the goodness of the result of the analyses. Img.Seg. is an operation of cutoff an image to various areas, such that each region is similar to else [7] [8]. The segmentation model for monocular images can be expanded to color segmentation by using "RBG" or their conversion (linear/non-linear). Although, global surveys on color image segmentation are few [9]. Analyzed the issue when applying edge and area segmentation mode to color images with synthesis texture. [10]. Segmentation is a pertinent technique in image treatment. Various methods occur in multi-apps. Histograms are found to be very efficient in terms of calculation complexity when compared with other division methods. If identified low, High point is properly and proper threshold is fixed, this technique will show good output results [11] It was also implemented in converting DICOM files into a three-dimensional model. [6].
It can take advantage of a set of elements with high-resolution features, such as pixels, texture, and shape properties. Threshold nodes a very simple algorithm to perform hashing. This threshold optimum can be calculated by simulating a Gaussian normal distribution Equation 1 Figure 1, of two regions of the image and calculating the middle and norm deviations of the picture through classified region interest, edges, and background. [12] [13].

Anatomical model & visualization system
A lot of applications working in surgical and medical image research called Anatomical models in the clinic's the body natural interaction with Virtual-Models simplify anatomy in training and different interactive structures spatially in the human body. Mixed reality is a part of a Visualization system called training Simulation with virtual models anatomy, where to reduces the surgical involvement which is linked to patient risk and cost of healthcare [14]. MR is part of the visualization systems, and it works to create a new reality by integrating a realistic environment with the virtual, which allows the integration of real models and virtual models, to enhance the positive indicators of the surgeon. Applied to liver surgeons during training, before, during, and after the surgical operation in terms of visualization techniques [15]. For example, using Mixed Reality (MR) is an advanced technology used to live in a real environment with fictional models, and it also helps the user to determine its location and medical model location information by Azure Spatial Anchors. Helps classify and store medical images by "Azure Cosmo DB", using Internet of Things services [16]. HOLOLENS devices can live in the real environment and recover 3D models by improving display speed during training in real-time, by increasing the accuracy of the intensity and degree of illumination while taking into account the degree of illumination distributed in the area to improve the display intensity [17].

CNN in medical segmentation
CNN: "Convolutional-Neural-Network", a part of ANN: "Artificial-Neural-Networks" has been prevailing in different computer vision objectives, benefits across a set of domains consist of radiology. CNN designed a spatial hierarchy of features through back-propagation out of a lot of building layers automatically [18]. CNN's hold three types of layers, A) convolution layers, where a kernel of weights is convolved to extract features; B) pooling layers, which replace a small district of a feature map with some statistical input (mean, max, etc.), and C) nonlinear layers, apply an activation duty on feature maps, to can the designing of non-linear duty by the network [19] [20]

Related Work
The researcher uses part of deep learning called a steady architecture CNN with input trip planar perpendicular patches while be used with advances in convolution kernels and deep networks The fulfillment of totally linked layers as 1×1 convolution layers and the skip from Shorthand layers pliable rapid diversion of full figures compared with more taking longer timing scanning Moreover Input 3D information patches force result in increased execution performance but require calculation the increased load of the parameter network [21] A new version method using deep learning techniques and dynamic random walker method has been proposed of MR brain image segmentation The hybrid method has shown that best execution as compared through another state methods [22]. The use of CNN and limitations of deep learning and all advantages is primary in a major of radiology research to improve performance radiologists and, patient care [23]. He used an automatic algorithm in liver segmentation to improve the accuracy of the segmentation problem, where it is based on combine between active area (contour) and collect data to use as input for deep learning networks [24]. Using the multi-view and the power of CNN to split, segment PPV and LA through cardiac-NET MRI. The method shown, a combination between different information inputs of MRI throw an adaptive combination planning and a new mission, improves damage segmentation and accuracy [25].
He worked on developing a framework to precisely segment all basic structures of the heart from calculating tomography and magnetic chest imaging together with high activity. Using multiple CNNs training from the start and permitting an adaptive information fusion strategy in the labeling of pixels was suggested despite data limitations. The results demonstrate a tool to accurately and efficiently identify cardiac structures called [MO-MP-CNN]. Instead of using 3D CNN which is a multi-level 2D CNNs. Thus, use a multi of 2D slides is more important than using a 3Dimentional model within the current setup. In the case of available a lot-of GPU processing and data 3D [26]. The author designed a framework called DMPCT for CT scan segmentation in multiorgan, which is Stimulate by the imitative co-exercise strategy to combine multi-planar input for the dummy data through training, without required a massive of 3D volumes labor from radiologists [27].

Proposed architecture
We propose a process to implement liver DICOM image segmentation, and use active contour technique as a method for segmentation as input data where use seg3D2 and CNN network of deep learning Figure 3. Next step ACM deduction, a structured see is calculated and taken to the CNN, which parameters can already exit updated using back-propagation.
Since a DICOM image is usually of low contrast, and it is difficult to distinguish between each element and the other and the background, the image must be processed first before starting to split and segment. Due to the large size of the image data and its lack of arrangement, the processing of images plays a large role in defining each element and clarifying the data. As it is the first and required step to improve the quality of the images to obtain high accuracy in reading the data. DICOM images consist of many noises and lighting impurities. Before starting to analyse, segment, and object detection, all impurities must be removed through the processing procedures first.
First of all, we must use the point of convergence and intersection of the DICOM files for the three images (three orthogonal) where (coronal=X, sagittal=Y, axial=Z =0) figure 4 as input data and compare them with the 3D model or real images to detect a liver object, Scientific Steps [1]. where the three images are the general shape of the model, and since the images are of low contrast, we should start Convolutional Layer then the histogram equation figure 5 for the raw images must be used as a technique for adjusting the image intensity to improve contrast as shown below figure 6.
The results show that the images with high noise use the Gaussian Blur technique, while the Gaussian function, in contrast, small and large. We, therefore, recommend Gaussian use, with alpha = 1.

Comparative process of medical image
Growing the number of point light [brightness and contrast] Fig.7 (b) to improve image Accuracy & density, as it becomes higher image quality to make it easier to convert into a 3D-model. In Scientific Steps [2] present image processing result of DICOM image before and after increasing point light on the image through using python and specializes in image-processing, data-visualize and data-analysis called open-CV and NumPy library, to ensure DICOM empty from impurities and prepared to Convert.
There are no differences in the details and content of the image before and after increasing the brightness and width ratio Fig. 7, the appearance of impurities only in the borders Fig. 7(C) of the image, but it does not affect the conversion process in the three-dimensional model. The impurities are removed and the model is purified with MESHMIXER.

Active Contour Model (ACM)
ACM is the practical framework [28] [29] of a two-dimensional potential image diagram called the snake model and is widely used in tracking objects, recognizing shapes, image segmentation, and defining edges. A true image, knowing that snakes do not solve the problem because of the clarity of the image's features, as the best way to know in advance the shape of the contour required so that the user can interact with the user with the highest-level images, or from all other information from the surrounding image data, time and place.
To obtain a good training result, the entire liver and its surrounding pixel units must be preserved, with the missing pixel versions eliminated and the liver frame deformed so as not to cause a bad prediction if versions of many good cases are available Figure  8.
ACM conclusion modified dependent on image and local point sentence equation terms, and expected loss that is used to train CNN to create these sanctions maps, the following Figure 8 is designed in Python to display the field of masks on the images. The proposed system of framework diagram is shown in Figure 3. http://www.i-joe.org

Medical 2D image structure generator
A single image is just a two-dimensional plane depicted through a projection of a three-dimensional object. Because some data is lost through the dimensions and area of the image that represent the lowest (2D) dimension from the top (3D), so no data available to originate a model-3Dimensional from 2Dimensional images Figure 9. But should generate a standard 2D CNN Structure include shape prior knowledge of the 3D object. from a single to multiple 2D-image projection mapping Figure 10

Results
This section shows the result of active contour images from the DICOM raw image to the 3D-model. First, we show the result of the pre-processing from DICOM raw images. Each image contains edges and image size =256 X 256, which through adjustment and control we can create multiple results for output using Active Contour, by resizing the threshold frames and gray-scale image, shown below in table 1. Second, we can generate a lot of versions of active contour 3D-model output result, by changing the Histogram and Gaussian equation, and many of transfer function. will show all functions as shown below "table 2". Now we can show 3D-model Deep architecture, it is through the CNN, through which the images of the three axes X, Y, and Z are determined and matched with a real image of the body shown above Figure 6, and the area required to be determined as shown above Figure 9 and edits the contrast using a histogram. Therefore, its prior knowledge and identification of the three dimensions of the images (three orthogonal) (coronal = X, sagittal = Y, axial = Z = 1) to intersect them at one point and determine the liver area and convert to the 3D shape model Figure 11, Figure 12.  To receive grate experience and training result, we should to keep versions of the active contour 3D-model which include full pixels knowledge of the liver and delete other versions which the lost pixel of a liver model, because other versions will be made more distortion when re-Sculpting & Trimming 3D-liver model by 3d programs Figure  13. This significant step will be to test the final simulation through a 3D-model and converted the OBJ file to a low poly object by the modeling process steps sculpt, smooth, and test the model motion in the graphics programs such as using "Adobe MESHMIXER". Therefore, in this status can be estimated outer perimeter shape stabilization and mass the liver and kidneys volume, wherever can evaluate the internal mass (Volume) by mm3 and the external perimeter (Surface Area) by mm2 Figure 14.
Now can be compared through the combination of hybrid algorithm structure (CNN, Active Contour & Deep Multi-Planar) and seg3d2, the use of Python programming language and open-source libraries OPENCV, CV2, and NUMPY, and real results through the liver volume analysis report from MeVis Distant Services (1731 ml), where the volume of the liver is measured in milliliters ML, and our results are MM3 (1.73059e+06 mm 3 )as shown Figure 14, so the final results were converted into milliliters ML to ensure the correct size of the liver as shown Figure 15.

Conclusion
We proposed in this research a simple efficient hybrid technique based on two algorithms "ACM and CNN", which has applied the power of deeply CNN to DICOM image and enhancement segmentation data files. The system achieved a high result of segmentation, each image contains edges and image size =256p X 256p, using Active Contour Model can generate multiple results for output, by resizing the threshold frames and gray-scale image and adjustment control, and generate a multi 3D model as output by changing the Histogram and Gaussian equation. In the future work, the proposed system could be utilized for detecting the Cancer and position accuracy in the Liver CT images.  EGYPT 20-9-1970) received B.Sc. degree in Electronics Engineering, and M.Sc. in Electrical Communication Engineering from the Faculty of Engineering, Mansoura University -Egypt, in 1992 and 1995 respectively. Dr. El-Bakry received Ph. D degree from University of Aizu-Japan in 2007. Currently, he is full professor at the Faculty of Computer Science and Information Systems -Mansoura University-Egypt. He is the head of Information Systems Dept. His research interests include neural networks, pattern recognition, image processing, biometrics, cooperative intelligent systems and electronic circuits. In these areas, he has published many papers in major international journals and refereed international conferences. According to academic measurements, now the total number of citations for his publications is 3698 and the H-index of his publications is 30. Dr. El-Bakry has the United States Patent No. 20060098887, 2006. Furthermore, he is associate editor for journal of computer science and network security (IJCSNS) and journal of convergence in information technology (JCIT). In addition, he is a referee for IEEE Transactions on Signal Processing, Journal of Applied Soft Computing, the International Journal of Machine Graphics and Vision, the International Journal of Computer Science and Network Security, Enformatika Journals, WSEAS Journals and many different international conferences organized by IEEE. Moreover, he has been awarded the Japanese Computer and Communication prize in April 2006 and the best paper prize in two conferences cited by ACM. He has also been awarded Mansoura university prize for scientific publication in 2010 and 2011. Dr. El-Bakry has been selected in who Asia 2006 and BIC 100 educators in Africa 2008. (Email: elbakry@mans.edu.eg).