Global Journal of Computer Science and Technology, C: Software & Data Engineering, Volume 22 Issue 2

Frameworks such as MapReduce [17] and Spark [40] are not suitable for models serving in real- time because they were designed and built for batch processing. Furthermore, they are not suitable for large numbers of small transactions because of the considerable time overhead that they require for instantiation. Dask [27] and Tensorflow [35] provide a complex and very little support for model serving [26]. It is possible to set up different parts of different frameworks together to have a system that can serve a deep learning model. However, the compatibility and maintenance of these different frameworks increase the technical complexity. Unfortunately, deploying deep learning models into production is still not a straightforward endeavor. b) Existing Microcirculation Analysis Systems This section presents the current work on systems that calculate capillary density from microcirculation images. As briefly mentioned at the end of the introduction, none of the existing works mentioned on microcirculation analysis reported using parallel frameworks to calculate the capillary Fig. 1: The diagram shows the code encapsulated in a core on a computer. This code is replicated across each core to achieve parallelism. This code calculates the capillary density from a microcirculation image. The architecture consists of two parts: i) first determining the RoI using traditional computer vision algorithms and ii) then using deep learning to classify if the RoI contains a capillary density in under 1 second for a frame with a resolution of 1920x1080 on a CPU using deep learning with an accuracy of 85%. Those who exceeded this accuracy used a GPU which is not readily available in a clinical environment. Cynthia Cheng et al. [41] takes a three-step approach to quantify capillary density. First, they apply an image enhancement process to darken the capillaries and lighten the background. They then flatten the image using 2D filters and raise the image’s contrast. The image is then despeckled using a 7x7 filter. They then adjust the histogram of the image to a best-fit model The second step involves manually selecting the capillary as a target object. They then select the background as a reference. The algorithm then selects the rest of the capillaries and excludes the images. A macro is then created from this process, which can be applied to other images with similar characteristics. As described, this involves several steps, including the manual user intervention; therefore cannot provide results in less than 1 second. A. Tama et al. [42] uses binarization followed by skeleton extraction and segmentation to quantify the capillaries. The first step involves extracting a reference image. The image has to be then manually cropped by the user. The green channel is then extracted from the image to have the highest probability of vessels in it. They then apply a top-hat transform to remove unevenness in the background. They then apply the Wiener filtering, a lowpass filter followed by Gaussian smoothing. They then apply Otsu thresholding to segment the image from the background and apply a skeleton extraction method to quantify the capillary. The authors do not report the speed needed to perform these steps. Sherry G.Clendenon et al. [43] uses a manual method to segment the microvascular structure. The authors do not report the speed or accuracy of their method. Pavle Prentaˇsic et al. [44] used a custom neural network to segment the foveal microvasculature. Their neural network consists of three Convolutional Neural Network (CNN) blocks coupled with max-pooling and a dropout layer followed by two dense layers. They reported accuracy of 82.4% at 2 minutes. R Nivedha et al. [45] used a non-linear Support Vector Machine [46] to classify images. They first started by extracting the green channel since it contains the Capillary X: A Software Design Pattern for Analyzing Medical Images in Real-time using Deep Learning Global Journal of Computer Science and Technology Volume XXII Issue II Version I 15 Year 2022 ( ) C © 2022 Global Journals ∼ ∼