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

(e) (f) Fig. 2: (a) This presents a sample of a microcirculation image that is taken as an input to the system (b) The background image calculated using a Gaussian Segmentation Algorithm (c) The segmented area formed by calculating the difference between the original image and the background (d) The Structural Similarity Index calculated from the original frame and the background image (e) The modified image with the capillary area highlighted in black encapsulated within the green bounding box (f) The original image with the capillary area highlighted in black encapsulated within the a bounding box applied the Adam optimizer and cross-entropy as loss metrics. They trained the NVIDIA GeForce GTX Titan X algorithm and used the PyTorch library. They reported accuracy of 88%. However, such an algorithm is not suitable for a clinical environment due to the high-end GPU required to run it. F Ye et al. [51] utilized the concept of transfer learning and used the Inception Single Shot Multibox Detector (SSD) [52] to build their neural network. They build their system using Python and Tensorflow with an image resolution of 744 × 482 pixels. They applied data augmentation to the image to increase the number of datasets. The SSD architecture requires GPU to produce results in real-time, making it unsuitable to be used in a clinical requirement with only CPUs available. YS Hariyani et al. [53] used U-net architecture combined with a dual attention module. They introduced a new method called DA-CapNet, which can analyze microcirculation images. It consists of the encoder and decoder parts. The encoder downsamples the dimension of the information in an image while increasing the number of channels. This step increases the spatial information dimension. They then combine it with a dual attention module which increases the accuracy. The dual attention uses the squeeze and excitation process to extract the blood vessels in the image. The authors resized the image to 256×256 to reduce the processing time and used a Gaussian threshold method with a median blurring filter of kernel size five. The authors reported accuracy of 64% but not the time taken for analyses. G Dai et al. [54] used a custom neural network similar to Pavle Prentaˇsic et al. for segmentation. However, G Dai et al. used five CNN blocks instead of three. Hang-Chan Jo et al. [55] used a Attention-UNet architecture [56]. Their method starts by using the CLAHE method and computes several histograms. They then apply the Gamma correction and pass it to the deep neural network. The reported accuracy was 73.20%, but not the time is taken for analysis. III. P roposed S ystem This section presents the system’s architecture to analyze medical images in parallel, specifically, to calculate the capillary density in a microcirculation image. We start by presenting the DL part (which is based on OpenCV [57] and Tensorflow [35]) and the architecture of our system’s parallel part (which is based on Ray [58]). a) The Deep Learning Algorithm part of the Proposed System The outline of the deep learning architecture is shown in Figure 1. It consists of two main parts: i) determining the regions of interest (RoIs) where capillaries might exist, and ii) using a CNN for predicting whether these RoIs contain a capillary or not. The original frame is shown in Figure 2a. The position of the capillaries is determined by first removing the background from the original frame using a Gaussian Mixture-based Background/Foreground Segmentation Algorithm [59]. The background removed is shown in Figure 2b. The structural similarity index measure (SSIM) [60], [61] is applied between the original frame shown in Figure 2a and background image shown in Figure 2b resulting in Figure 2c and Figure 2d. Bounding boxes are formed around the red areas using OpenCV contour method [62]. These bounding boxes are then passed to the CNN for prediction. The RoIs that have been predicted as capillaries have a green bounding box around each one of them along with a black line to highlight the shape of 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 17 Year 2022 ( ) C © 2022 Global Journals