Performance Analysis of Convolutional Neural Network for Pneumonia Detection on an Embedded AI System

The increasing use of Artificial Intelligence (AI) in various fields has increased the need for a large amount of data. A device with adequate computational power is required to manage the data and produce an output with high processing speed and satisfactory accuracy. Meanwhile, the use of several embedded system devices for neural networks (NN) is constrained by low processor and memory capacity. Several embedded-system devices with improved processor capabilities have been developed for NN data processing. This study analyzed the capabilities of an embedded system device for NN in health applications, namely, the detection of X-ray images of patients with pneumonia using a Convolutional Neural Network (CNN). 2D CNN architectures have been employed with various parameters, including color depth, layers, filters, kernels, and quantization. The outcome was expressed in terms of the accuracy, inference time, RAM, and Flash consumption. The results showed a significant positive association between all output metrics and the number of filters. However, in some situations, the RAM and Flash utilization of the embedded system exceeds its capacity, making it unusable. This indicates that the inference time and memory are influenced by the number of layers, filters and quantization.