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

values were calculated by measuring the time to process a frame in a set of 100 microcirculation images. The execution time of the three architectures is presented in Figure 5. In short, our results show that our proposed system is 12% faster than the baseline parallel architecture and 78% faster than its baseline serial architecture. b) Speedup This metric calculates the speed gain by the system as the number of cores increases. For the baseline serial architecture, the execution time is one second regardless of the number of cores available (indicating that the system is not scalable). The average execution time of processing one frame for the baseline parallel system and the proposed system is shown in Figure 6. Fig. 7: This graph shows the number of cores used by each architecture to process a frame. The more used at any instance the better since this shows how efficient the system is at utilizing all the resources available to it. The baseline parallel system processed a frame on average in 0.56 seconds with four cores, while the proposed system processed a frame in 0.32 seconds. The proposed system processes a frame 68% faster than the baseline serial architecture and 43% faster than the baseline parallel architecture. As the number of cores doubles, the proposed system gains an additional 31% during the baseline parallel architecture gains an additional 55%. In both cases, the proposed system outperforms the baseline parallel architecture. One of the main reasons the proposed system outperformed a baseline parallel system with a master-slave architecture is that a masterslave architecture can reserve up to two cores to manage the other parallel cores. In contrast, the proposed system does not reserve any cores beforehand. In this way, we free up the computer cores to focus on processing images rather than purely handling requests. Thus, the proposed system gains more speedup than the baseline parallel system. We can conclude that the proposed system has the recommended architecture for running deep neural networks on a single machine. c) System Resource Utilization - CPU Usage This metric measures the number of cores used to process the medical images using deep learning. With the baseline serial architecture, only one cores is utilized per frame due to the Python Global Interpreter Lock’s limitation. With the baseline parallel architecture, it is always two less than the available number of cores because it always reserves these two for the management of the parallel workers. Each cores in the proposed system is allocated a task where each task processes a frame. Thus, the proposed system is most efficient on a single machine with a multi-core. A graph showing the number of cores used by each architecture is shown in Figure 7. d) System Generalization Our system functions and classes were built using modular design patterns. This design philosophy means that the user can replace the DL part of our system with their algorithm by simply pointing the function in our code to their algorithm. The details of this are highlighted in the README file in the GitHub repository. Thus, our package can be generalized to analyze images using a DL model of the user’s choice in parallel. The system will automatically scale to the number of cores available without the user having to worry about experiencing issues with dependency, integration, resource utilization, and speedup. VI. C onclusion This paper presented a software package that can analyze medical images using DL locally. Our proposed system can efficiently use all local resources because it utilizes parallel execution to offset the resource-intensive demands of using a deep neural network. The proposed system is of high clinical relevance because monitoring changes in capillary density can be used to locate early markers indicating organ failure. The severity of the change in capillary density might predict whether or not the patient survives. Furthermore, clinical researchers do not risk uploading patient data to a third-party cloud provider to use a deep neural network to automatically analyze their images. ∼ ∼ ∼ 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 20 Year 2022 ( ) C © 2022 Global Journals Fig. 6: This graph shows the speedup of the baseline parallel system and the proposed system as the number of cores increases