Pneumonia is a lung infection caused by bacteria, viruses, mycobacteria or fungi [1]. Common symptoms found in pneumonia patients are coughing, fever, shortness of breath, nausea, and vomiting [2]. Pneumonia diagnosis is made by listening to the sound of breathing, which is usually heavy and rumbling. The next step is to use X-ray technology to capture photographs of the chest region to observe the lung area. The location of the infection in the lungs is determined from the collected images to determine the area and degree of dissemination. Additionally, a patient's blood is typically examined in a laboratory to establish the cause.

One of the main causes of pneumonia is pollution [3], and in recent years, pneumonia has been caused by the COronaVIrus Disease of 2019 (COVID-19 )virus [4]. Pneumonia is a malignant disease that can lead to death. Pneumonia patients range in age from children to elderly individuals. In developing countries, patients with pneumonia do not receive effective treatment. The lack of effective treatment potentially contributes to the high mortality rate associated with this disease [5]. However, in developing countries, diagnostic equipment is expensive. Devices with high computational capabilities are available only in large hospitals in large cities. Therefore, a tool to assist doctors in detecting pneumonia that is inexpensive, does not require complex treatment, can be used easily, and allows the device to be used as a data source, which is the core concept of edge computing (EC), is needed.

Pneumonia has been identified in multiple studies employing patient X-ray images and various methodologies. Several image-processing techniques have been applied, such as integrated filters, enhanced contrast, edge detection [6], and texture analysis [7]. Another approach is to use the Earth Mover’s Distance (EMD) to identify the infected lungs and normal lungs using the distance between two probability distributions [8]. Furthermore, to assist this analysis with a computer-aided diagnosis system, several studies have been conducted, including machine learning (ML) [9], deep learning (DL) [10]–[13] and ensemble DL [8], [12], [14], to improve the performance of CNNs. This ensemble DL combines several well-known CNN model outputs during the experimental phase. Another CNN technique based on semantic and instance segmentation, called the mask region-based convolutional neural network (Mask RCNN), was applied to pneumonia X-ray images [15]. However, all these CNN approaches require prediction, detection, and classification. On the hardware side, this condition calls for additional components, such as the graphical processing unit (GPU) processor, to accelerate and alter images in memory [16], [17]. In this study, computing was performed on computer, and the deployment of DL to an embedded-system device that does not have GPUs.

Several previous studies have used microcontrollers for machine learning (ML). A support vector machine (SVM), k-nearest neighbor (kNN), and decision tree were deployed to the STM32 Nucleo Board [18]. To provide an automatic intelligent system, kNN was applied to the TI CC1110F32 microcontroller [19]. However, running DL on embedded systems that have limited capacity is still complicated because DL applications require quick and precise data processing owing to the vast volume of input data [20]. The physical size [21], data processing speed [22], memory usage [23], and electric power consumption [24] are important factors in the development of embedded systems for DL. Several efforts have been made to make complex calculations more efficient and reduce memory usage in DL by making adjustments to the DL architecture [25]–[28] or changing variable parameters from 32-bit floats to lower parameter types, also known as quantization [29], [29], [30].

Currently, sizable data are commonly sent from edge devices and processed in the cloud. The results are then sent back to the edge. However, this process is time-consuming. Therefore, a device that performs processing directly on the edge is currently under development, which would save time [31]. This study offers a further option for CNN applications in embedded systems by performing a variety of analyses of pneumonia X-ray images on a number of input and output characteristics to determine whether they can be used in embedded systems designed specifically for AI. This analysis was also intended to provide a general overview of the issues that may arise when a CNN is implemented in an embedded system. The selected application was a medical-related application because current medical devices are required to be portable devices that can be used anywhere.

Numerous output measures were obtained from the training outcomes. The output can also be displayed as a number using a 2-dimensional visual graph, as illustrated in Figure 5. The figure shows the data that were properly guessed as well as the wrongly guessed data. This visualization makes the analysis easier by providing a broad perspective of the position of the data, if shown on a graph.

Figure 5 Figure 5 Examples of visualization of the full training dataset: (a) 16/32, 2 layers, grayscale, 3 layers, grayscale, kernel size of 3; (b) 16/32, 2 layers, RGB, kernel size of 3 Source: Own elaboration.

The CNN architecture, which consists of two layers and three kernels with grayscale color depth, and tests with various filter pairings are presented in Figure 6. The findings demonstrate that increasing the number of filters in each layer improved the accuracy from 86.1% to 94.8%. Moreover, the inference time exhibited exponential growth. This demonstrates that as the number of filters increases, the computational complexity of the CNN architecture likewise increases, thereby lengthening the inference time. The RAM and flash usage also increased as the number of filters increased for int 8, as shown in Figure 6a. The inference time significantly increased when the number was changed to 32, as shown in Figure 6b. Additionally, it results in a large increase in RAM and flash memory. This reveals that filter 32/64 for both types of numbers and 16/32 for float 32 use more RAM than does 256 K, which means that they cannot be deployed to NANO BLE 33.

Figure 6 Figure 6 2D CNN, grayscale, 2 layers, kernel size of 3 (a) int 8 and (b) float 32 Source: Own elaboration.

The RGB color depth, which has three times as many features as grayscale features do, yields nearly the same results, as shown in Figure 7. The accuracy, inference time, RAM, and flash utilization increased as the number of filters increased. The inference time, RAM, and flash utilization were clearly greater than the grayscale color depth. This is understandable because of the large number of inputs. However, there is no change in the accuracy between the number types int 8 and float 32. The RAM utilization in the 4/8 filter for float 32 is quite close to the NANO BLE 33 RAM's maximum capacity; therefore, it is not advised to use it in this situation according to the RGB input. It is clear from the additional filters in this section that more RAM is needed than the amount of RAM on the NANO BLE 33.

Figure 7 Figure 7 2D CNN, RGB, 2 layers, kernel size of 3 (a) int 8 (it demonstrates that because filter 32/64 uses more RAM than 256K does, and it cannot be deployed to NANO BLE 33) (b) float 32. (It shows that in all cases, this model cannot be deployed since RAM usage is more than 256K) Source: Own elaboration.

The addition of one layer to the architecture with color depth grayscale also increases the accuracy and other output parameters, as shown in Figure 8. There is an insignificant increase and even a decrease in accuracy when layers are supplemented with a 128 filter. The other output parameters were also increased. Compared with the grayscale architecture using two layers, the inference time is longer. However, the use of RAM and flash memory is similar. There is a significant difference in the inference time, RAM and flash usage when quantization is used without optimization. Similar to adding a layer to the grayscale input, this is seen in various situations but cannot be used with NANO BLE 33 because it uses more RAM than is provided. For example, on float number 32 for type int 8, filter 32/64/128 and all other filters aside from filter 4/8/16.

Figure 8 Figure 8 The 2D CNN, grayscale, 3 layers, and kernel size of 3 demonstrates that because the 32/64 filter uses more RAM than the 256K filter does, it cannot be deployed to NANO BLE 33 (a) int 8 (b) float 32 Source: Own elaboration.

Figure 9 shows the output using the RGB input and three layers. The same tendency, that is, an increase in all parameters, was also observed when the RGB input was used. In contrast to the grayscale input for quantization, this had a smaller impact. With the same number of screens and grayscale inputs, it is clear from the unoptimized number that the inference time varies significantly. However, compared with the grayscale input with the same number of layers, the RAM and flash utilization did not noticeably change. The RGB input and the three layers exhibit nearly identical patterns for RAM utilization that exceeds the RAM capacity and cannot be used with NANO BLE 33.

Figure 9 Figure 9 The 2D CNN, RGB, 3 layers, and kernel size of 3 demonstrates that because the 32/64 filter uses more RAM than the 256K filter does, it cannot be deployed to NANO BLE 33 (a) int 8 (b) float 32 Source: Own elaboration.

The accuracy and use of RAM and flash memory improved little or not at all after the kernel size was increased. There was an even decrease in the accuracy when the kernel size was 4. However, the inference time appears to have significantly increased for both the quantized and unoptimized values, as shown in Figures 10 and 11. The use of a kernel size of 3 resulted in the best performance. Additionally, the use of RAM has no effect on the type of number without optimization (float 32).

Figure 10 Figure 10 Various kernel sizes of 2D 16/32 filter CNN, grayscale, 2 layers (a) int 8 (b) float 32 Source: Own elaboration.+

Figure 11 Figure 11 Various kernel sizes of 2D of 16/32 filter CNN, RGB, 2 layers (a) int 8 (b) float 32 Source: Own elaboration.

Reducing inference time and memory usage is essential for embedded AI system applications, such as agricultural and health surveillance edge devices placed in resource-constrained areas, to improve energy usage and increase battery life. Therefore, by adjusting the hyperparameters on the basis of the needs of a given task, resource usage can be optimized without reducing the performance of the device.

To determine the significance of the correlation between the output parameters, a correlation test was performed using Pearson correlation, and the results are shown in Table 2. The table shows that the correlations between accuracy and inference time and between RAM and flash usage are positive in the range of 0.5–0.9, which means that an increase in the number of filters on the layer increases accuracy, indicating a very strong positive correlation with other parameters.

The lowest correlation is seen in the use of accuracy vs. the use of flash on grayscale, type number int 8, and layer 3 (e.g., r=0.558671), and the highest correlation is seen on grayscale, type number float 32, and three layers (e.g., r=0.928159). In layer 2, the correlation results of the 8-bit quantized numbers and 32-bit unoptimized numbers are not significantly different. In layer 3, there is a significant difference between 32-bit unoptimized and 8-bit quantized numbers. However, the use of the 32-bit unoptimized model also results in fairly large memory usage on embedded systems; therefore, it is necessary to consider its use on embedded systems that have a small memory capacity. In this case, the use of an 8-bit model suggests high performance at low computational and energy costs.

Table 2 Table 2 Correlation of increasing filter number with various parameters using Pearson correlation Source: Own elaboration.

Previous research has also examined the application of medical image recognition in embedded systems using CNNs, which also reveals that computational and memory limitations should be considered, as shown in Table 3. The use of embedded systems has been studied by researchers using CNN architectures or models specifically designed to be deployed in embedded systems with limited capacity [34]. Moreover, the use of embedded systems such as minicomputers results in relatively fast inference times in medical applications [35], but it consumes more RAM. Another approach is to use a special random-access memory (RAM) architecture for a CNN to improve energy efficiency [36]. As an alternative, a field programmable gate array (FPGA) [37] has been used to perform specific classification tasks using CNNs with the aim of optimizing image augmentation and preprocessing data with hardware limitations. The inference time was significantly faster, and the memory usage was relatively small. However, the use of FPGA requires special programming skills.

Table 3 Table 3 Comparison of CNN implementation on medical devices Source: Own elaboration.

Currently, several microcontroller devices can be used for relatively uncomplex AI applications at relatively low cost. However, for relatively complex health applications that require high accuracy, minicomputers and FPGA, which are relatively expensive, are still required. Another critical factor is the availability of electricity in remote areas or whether the device is placed as a monitoring device. The solution that can be provided is to use solar cells or other renewable energy sources, such as wind. Therefore, a quantized model was developed to reduce the electricity consumption of the devices. Likewise, devices must be equipped with an intuitive interface, minimal maintenance and setup requirements, and a sturdy and robust casing to protect the device from various conditions. Devices should also be equipped with a manual in the local language so that it can be understood by users.

For future work, further research can be performed using advanced quantization techniques, such as mixed-precision [38], to improve accuracy with low memory usage. Another alternative is to use network pruning techniques [39] by eliminating redundant filters. Future research can also be conducted utilizing a more intricate multiclass classification, various NN architectures, or a moving object to determine whether embedded systems are capable of performing calculations with a higher level of complexity. In addition, real-world validation is needed for various applications in the medical field. This research can contribute to the creation of inexpensive, portable and intelligent medical equipment.

In this study, the application of DL for pneumonia detection was analyzed using X-ray images from an AI-based embedded system. Several input parameters, such as the number of layers, filters, kernels and input features, were varied to determine how these changes would affect the output parameters, such as accuracy, inference time, and RAM and flash usage. This was achieved by considering the limited memory and microprocessor capabilities of the embedded systems.

The results showed that, in some cases, an increase in filter number can increase accuracy, but this comes at the cost of increased inference time, RAM and flash usage. Therefore, it is necessary to focus on the correct configuration of the input parameters according to the capabilities of the embedded system without neglecting the output parameters. The use of the 32-bit unoptimized model also results in fairly large memory usage on embedded systems, so it is necessary to consider its use on embedded systems that have a small memory capacity.

In medical applications that require real-time and efficient decision-making, accuracy and inference time are critical. The two variables must be considered when selecting and determining a model. This is because medical applications are highly sensitive to various environmental conditions. Model optimization with quantization or other techniques is essential for creating a precise AI-based pneumonia detection tool for use in low-resource clinical environments.

On behalf of all authors, the corresponding author states that there are no conflicts of interest.