An artificial intelligence-enabled consumables tracking system for medical laboratories

2.1 Barcode technology

In the period before the implementation of the laboratory management and tracking system, the handling of consumables primarily relied on manual processes. Handwritten annotations on paper played an important role in the original process. This approach required the personnel to manually transcribe inventory details [29]. However, this conventional methodology was susceptible to errors due to the potential illegibility of handwriting, resulting in misinterpretation or the incorrect placement of crucial information [30]. Moreover, this manual process frequently gave rise to challenges such as unforeseen product stockouts or excessive inventory, leading to inaccurate quantity calculations.

Over the past two decades, with the rapid advancements in computer technology during the digital age, a paradigm shift has occurred in the manual handling of consumables in the medical laboratory system. This transformation has effectively addressed the previous challenges and is now extensively employed in database management [31]. The integration of various technologies has significantly streamlined the management of diverse materials, with barcode technology standing out as a notable example. This technology has now become an essential element within contemporary modern information management systems in numerous industries and has revolutionized monitoring, identification, and data collection processes [32,33]. The fundamental concept that underlies barcode technology is the encoding of information through a visually scannable pattern consisting of parallel lines and gaps. This pattern can be quickly interpreted by specialized scanners or readers [34]. The technology requires only a few supplementary tools but has revolutionized supply management by reshaping how inventory is monitored and regulated. The roots of barcode technology can be traced back to the 1970s with the introduction of the Universal Product Code barcode [35]. Subsequently, barcode technology has undergone a substantial evolution, marked by the incorporation of advanced symbology like QR codes and the enhancement of barcode scanning capabilities. In contrast, QR codes, which were developed during the late 1990s, are matrix codes that include a two-dimensional structure, allowing for the storage of data in both horizontal and vertical orientations within a grid composed of black and white squares and thereby offering greater data capacity and versatility [36]. This technological progression has profoundly impacted consumables management by providing a distinct identifier for each item, facilitating efficient data collection and precise stock level tracking [37].

2.2 Radio frequency identification (RFID)

Barcodes have had widespread usage across diverse applications due to their cost-effectiveness, convenience of use, and efficacy for data encoding and identification. However, this technique still has a few restrictions. These restrictions result from the two-dimensional, static nature of barcodes, which limits the amount of information they can store and necessitates scanning in a direct line of sight [38]. Previous studies also reported that barcode scanning fails to achieve the intended error prevention results; therefore, some hospital staff have perceived barcode scanning to be ineffective for healthcare or job performance [39,40,41]. To address these limitations, the emergence of RFID technology has led to simpler and more efficient methods with which to accomplish comparable objectives. Barcodes effectively fulfill the role of product identification and inventory tracking; however, they necessitate line-of-sight scanning and lack the capability to scan multiple items simultaneously [38]. Conversely, RFID technology offers a multitude of advantages, notably the ability to encompass wireless data transmission and non-line-of-sight scanning and the capacity to read multiple tags simultaneously. Enabled by radio waves, RFID tags are effortlessly readable, facilitating swift and precise data recording [42]. In addition, the implementation of RFID technology eliminates the necessity for manual scanning, resulting in a reduction in human error and an enhancement of operational efficiency [43]. This characteristic contributes to improved supply chain visibility, greater inventory management, and increased overall efficiency. RFID technology represents a significant breakthrough in the streamlining and improving procedures that previously depended on barcodes. Despite the numerous benefits of RFID systems in consumables management, limitations remain. The difficulty involved in accurately tracing products under specific temperature conditions, especially extreme heat or cold, is a significant disadvantage [44,45]. These conditions can compromise the performance of RFID devices, leading to inaccurate readings, unreliability, or even complete failure [44,45,46]. In industries such as those dealing with pharmaceuticals and perishable products, where temperature control is essential, this limitation has significant repercussions. Inventory tracking based solely on RFID technology may not provide the required level of precision and confidence. To guarantee the accurate management of consumables, the incorporation of alternative solutions or additional monitoring techniques becomes essential.

2.3 ML and AI

In recent years, there has been a significant advancement in AI technology, particularly in the field of DL. DL has demonstrated the ability to effectively store and acquire certain traits by utilizing neural networks with diverse topologies that are adapted to their unique qualities. This technology emerged as a dominant field in the early 2000s subsequent to the rise in popularity of support vector machines, the multilayer perceptron, artificial neural networks, and other neural networks with fewer layers. Numerous researchers have classified DL as a subset of ML, which is itself a subset of AI [30,47,48]. Since 2006, this algorithm has seen a significant transformation and has gained considerable popularity in comparison to other ML algorithms of its time [49]. This has been attributed to two primary factors: the substantial volume of data for processing and the ready availability of advanced computational resources.

The field of object detection has recently seen significant advancements in DL, culminating in the emergence of the you only look once (YOLO) framework. In its earlier stages, object detection was heavily reliant on traditional computer vision methodology. The development of DL and, more specifically, convolutional neural networks (CNNs) made a significant contribution to the progress made in the field of image classification. The advancements in object detection, particularly R-CNN, Fast R-CNN, and Faster R-CNN, have played a crucial role in improving the efficiency and precision of this procedure [50]. These advancements were made possible by the integration of regional proposal networks. The YOLO algorithm, which was first introduced in 2016, brought about a significant transformation in the field by presenting a comprehensive approach that enabled the simultaneous prediction of object bounding boxes and class probabilities in a single iteration through the network [51]. Subsequent modifications, including YOLOv2, YOLOv3, and YOLOv8, have subsequently improved upon this groundbreaking foundation [51,52,53]. Several approaches had been employed to extract information from the images of the medical laboratory report, such as table and textual information [27,54]. In addition to addressing the support workload for diagnosis in medical laboratories, previous studies have employed this approach for the automated detection and counting of blood cells from smear images, achieving a processing time of less than one second [55]. However, the existing literature has not considered this method to improve the management system for consumable products in medical laboratories. To alleviate the staff’s workload by streamlining various operational processes and to improve the efficiency of medical laboratory management, an accurate and reliable tracking system for consumable products is urgently required.

Generally, DL-based object detection can be broadly classified into two distinct categories: anchor-based detectors and anchor-free detectors [56]. To improve the prediction of bounding box locations with additional offset regression, anchor-based detectors are constructed based on the concepts from traditional sliding window detection procedures. These anchor boxes are then classified as positive or negative patches and subsequently utilized to enhance the accuracy of bounding box predictions through the application of additional offset regression techniques [57]. These detectors, such as Faster R-CNN, Single Shot MultiBox Detector (SSD), and RetinaNet, exhibit a shared fundamental basis. Faster R-CNN achieves accuracy via the region proposal network (RPN); however, it demands significant computational resources and is designed to achieve real-time performance by utilizing a dense set of default boxes [58]. However, it may have limitations in accurately detecting small items. Meanwhile, RetinaNet addresses the issue of class imbalance with its focusing loss technique, but it encounters difficulties with densely overlapping objects [59]. On the other hand, anchor-free detectors signify a significant departure from traditional object identification methods by eliminating the need for predetermined anchor boxes. This approach provides simplified yet effective solutions for object detection tasks [57]. Among the various inventive ideas, YOLO emerges as a prominent example. One notable feature of the YOLO algorithm is its single-stage architecture, which facilitates real-time object recognition without the requirement of elaborate post-processing procedures [59]. Furthermore, the YOLO algorithm has strong adaptability to objects of different scales, and it exhibits reduced susceptibility to the anchor configuration difficulties encountered by anchor-based detectors [57,58,59].

In terms of application in the medical laboratory, when comparing YOLO to other object detection models, such as Faster R-CNN, SSD, and RetinaNet, several key advantages become apparent. First, YOLO excels in speed and is capable of processing images at an impressive rate of up to 30 frames per second or more while maintaining high accuracy [60,61]. Previous studies have shown that, in the application of object detection for real-time pill identification in a hospital pharmacy, YOLOv3 had better detection performance in terms of speed compared to the RetinaNet and SSD [28]. Its real-time capabilities make it suitable for tasks like real-time object detection in videos or cameras, which is particularly valuable for medical laboratory inventory management [62], enabling rapid sample analysis for timely disease diagnosis [63,64]. Additionally, the ability of YOLO to detect multiple objects within a single image is significant, particularly in scenarios with overlapping objects, such as in the simultaneous identification of medical waste materials in a laboratory setting. Furthermore, YOLO demonstrates precision in detecting small-sized objects, a crucial feature for research involving small-scale materials and details, such as text and numbers on medication labels. Another advantage is that YOLO eliminates the need for an RPN, which is required by other models like Faster R-CNN, resulting in a more accurate and faster detection process [53,61]. Lastly, the adaptability of YOLO with regard to variable image sizes enhances its utility in situations where objects and images come in diverse sizes [65].

In 2020, YOLOv5 emerged as the latest iteration of the YOLO object detection model; YOLOv5 was the most effective of all the YOLO iterations [66]. Other versions, such as YOLOv6, YOLOvx, PP-YOLO, and YOLOv7, which were variations and modifications of the YOLOv5 model, emerged around the globe [51,67,68]. In medical laboratory studies, it has been shown that YOLOv5 achieves the highest accuracy compared to the other models [65,69]. A recent study compared the performance of YOLOv5 and YOLOv7 in detecting gloves, masks, goggles, and lab coats in the laboratory monitoring system [65]. Although the study findings revealed that both YOLOv5 and YOLOv7 demonstrated strong performances, the study still discussed the fact that YOLOv5 performed better in the detection of small-scale objects [65]. Another study has highlighted the fact that the YOLOv5x model outperforms other sub-models within YOLOv5 in terms of average classification precision for object detection in pharmaceutical packaging [69]. The research compared the performances of sub-models within YOLOv5 in pharmaceutical packaging object detection and found that the YOLOv5x model achieved the highest average classification precision among its counterparts, including YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x [69]. In January 2023, YOLOv8, the most recent addition to the YOLO series, was released by Ultralytics; the same team released YOLOv5 in 2020 [70]. At the time of writing, it is the most sophisticated model and has improved feature aggregation and a mish activation function that enhances detection accuracy and processing speed [71]. However, only limited studies have investigated YOLOv8 in the medical laboratory. While YOLOv8 shows promise with its improved features, including enhanced feature aggregation and a new mish activation function, it is important to note that few studies have examined its performance in the context of medical laboratory applications, suggesting a potential area for future research and exploration [72].

Therefore, the objectives of the study are as follows: first, a complex DL framework within an infrastructure of tracking systems for consumable product tracking within a medical laboratory is designed and implemented. The goal of this objective is to modernize the laboratory systems for consumables management. Second, the performance of the model in identifying consumables is examined and validated by our research. We aimed to demonstrate the model’s proficiency with its fundamental tasks. It offers insightful information on the rapidly developing field of computer vision. Finally, our study conducts a comparative analysis of appropriate models for system integration between YOLOv5 and YOLOv8. It includes a thorough analysis of different DL models. In the end, the management system implements the integration of these models. The foundation of our research consists of these three interconnected goals. Through the lens of cutting-edge technology and thorough evaluation, it seeks to advance the field of consumables management within the context of medical laboratories.

3.1 Functional design of system

Our designed system belongs to the category of Laboratory Information and Management System, with a specific design for tracking consumables in medical laboratories. Utilizing AI, the proposed system aims to elevate the efficiency and accuracy in monitoring and managing consumable resources within the context of medical laboratory settings. In this study, we utilized the programming language C# in combination with the computer vision library OpenCV to develop a novel AI-based system for the purpose of monitoring inventory in medical laboratories. We developed our software by utilizing the C# programming language within the.NET Framework and employed this framework to integrate with the object detection model. The present system successfully integrates the programming capabilities inherent in C# with the image processing functionalities offered by OpenCV, thereby establishing a sophisticated platform. The designed system that was developed is illustrated in Figure 1, which provides an overview of the steps that are involved.

Figure 1

The architecture of the proposed system.

To begin, select a pre-trained object detection model that aligns with DL models. Following that, set up the C# environment by installing OpenCV libraries and the tools for image processing, thereby enabling effective collaboration with the DL frameworks. Once the environment is set, load the pre-trained object detection model within C# using the relevant framework’s C# API. Prior to inputting them into the model, ensure that the input images are preprocessed by performing tasks such as image resizing, pixel value normalization, and conversion to the model’s expected format. As the preprocessing stage concludes, execute the pre-trained model on the preprocessed images and effectively conduct object detection to yield valuable results, including bounding boxes and class probabilities for each detected object. Moving forward, implement post-processing techniques, such as non-maximum suppression, to meticulously eliminate duplicated or overlapping detections, while also filtering out low-confidence predictions. This organized post-processing aims to significantly enhance the overall accuracy of the detection results. To complete the process, craft compelling visualizations by adeptly adding bounding boxes and class labels to the original images, thus providing viewers with a clear and intuitive representation of the identified objects.

3.2 Material

The present study encompassed the utilization of a total of 30 consumable items in the course of conducting medical laboratory procedures. The inventory consisted of a total of 15 reagent bottles, 10 reagent boxes, and 5 distinct item boxes. The items discovered in medical laboratories were inherently unique, exhibiting variations in terms of style, size, color, brand, and other pertinent details. The study utilized the Omron FZ-SC Vision Camera for the purpose of capturing test data images. The camera under consideration possesses a pixel resolution of 640 (H) × 480 (V) and exhibits the ability to capture images encompassing a range of 12–480 lines. The camera employs a 1/3-inch image sensor.

The object detection models utilized for the identification of the consumable items mentioned in the study were YOLOv5 and YOLOv8. To optimize data processing capabilities, a Lenovo Think Centre computer equipped with an Intel Core i7-8700 processor, 32 GB of RAM, and a 256 GB SSD was employed. The aforementioned computer possessed the requisite computational capabilities and storage capacity to facilitate uninterrupted data processing.

3.3 Dataset collection and preparation

The images of the consumable items were obtained by capturing a total of 570 photographs using a vision camera; examples of these images are shown in Figure 2. These photographs were taken for a set of 30 items, with an equal number of images captured for each item on average. Every object was captured through the lens of a camera from various perspectives, under different lighting conditions, and with variations in the item itself. A stationary shooting position was established at a predetermined distance. The dataset was subsequently divided into three distinct sections, namely the training subset, the validation subset, and the testing subset. The training set comprised 15 images per item; the validation set comprised 2 images per item; and the test set comprised 2 images per item. This division ensured that the sets were characterized by a representative distribution of consumable items.

Figure 2

Representative dataset images.

The methodology for labeling these items with bounding boxes involves a systematic approach to ensure accuracy and consistency. Each dataset image containing multiple items undergoes careful annotation where every individual object is outlined with a bounding box. This process includes identifying and annotating each object separately to prevent oversight or inaccuracies, placing bounding boxes tightly around each object to accurately delineate their boundaries, and maintaining consistency across all images to uphold dataset uniformity and reliability.

3.4 Detection model design and training

The configuration process for detecting consumable items involved the utilization of YOLO models, specifically YOLOv5x6 and YOLOv8l. The annotated dataset was initially partitioned into two subsets, namely the training subset and the testing subset. Following this, a configuration file for the YOLO model was created; it outlined the network architecture and hyperparameters, such as the learning rate, batch size, and optimizer. This configuration file is crucial to the enhancement of the performance of the model. The model was trained using a dataset that had been annotated, and during this training process, the weights of the model were adjusted to minimize the discrepancy between the predicted bounding boxes and the ground truth bounding boxes. The hyperparameters were adjusted to optimize the performance of the model. In this study, we also implemented the early stopping mechanism with the models. This is a technique used to prevent overfitting to optimize the model’s performance [73]. It involves monitoring the model’s performance on a separate validation dataset during training. During training, it involves the monitoring of the model’s performance on a validation dataset. If the model’s performance on the validation dataset does not improve, or if it degrades after a predetermined number of training epochs, the training process is terminated early. This prevents the model from becoming excessively specialized to the training data and improves generalization to unseen data. This technique ensures that the model achieves its optimum performance while avoiding overfitting.

In the initial phase, predictions were made using visual representations of the consumable items employed in the training process. The results are presented in the form of graphs, specifically metrics/precision and metrics/recall graphs. These graphs depict percentage values ranging from 0 to 1 on the y-axis, while the x-axis represents the number of training iterations. The second phase entailed the assessment of the DL model through the utilization of a test subset comprising randomly chosen images, which served as a representative sample of the untrained objects. The purpose of this evaluation was to analyze the model’s capacity to effectively manage unfamiliar objects. This evaluation served as a supplementary metric for assessing the model’s performance, as depicted in Figure 3.

Figure 3

Training flow of the proposed system.

3.5 Assessment metrics

During the assessment phase, the efficacy of the DL model, denoted by equations (1)–(4), was measured using several metrics, including precision, recall, and mAP. In the realm of object detection, the intersection over union (IoU) serves as a fundamental metric, representing the intersection ratio between the model’s predicted bounding box and the actual data. These diverse metrics collectively offer comprehensive insights into evaluating our model’s accuracy across both object localization and classification tasks. By examining precision and recall alongside the extensively employed IoU and mAP, our assessment provides a refined insight into the model’s performance, going beyond the confines of traditional accuracy measurement.

In particular, IoU performs a crucial role in object detection by quantifying the overlap between the predicted and actual bounding boxes. This metric, defined as the intersection ratio between the two, serves as a quantitative measure of the model’s spatial alignment with ground truth data, adding granularity to our evaluation; this is defined as

In this context, A and B represent the areas of the bounding box and the predicted ground truth, respectively. When the IoU exceeds a given threshold, the detection is considered to be accurate.

Here, TP denotes accurately identified objects, FP represents objects the model incorrectly identified as positive detections, and FN refers to the entities that were not identified by the model, despite being present in the ground truth annotations.

The area under the precision–recall curve is denoted by AP. The mean average precision (mAP) is computed by determining the AP for each class and then calculating the class-wide mean.

To select the best model, the validation evaluation metric is mAP@(0.50:0.95), and we provided both mAP@(0.50) and mAP@(0.95) results on the test set. The mAP@(0.50:0.95) indicates the average mAP across various IoU thresholds (from 0.5 to 0.95 in increments of 0.05), whereas the mAP@(0.50) represents the average mAP over 0.5. To gain a more comprehensive understanding of the model’s performance across various categories of consumable items, statistical analysis and visualization techniques were employed.

4.1 Performance metrics of the proposed system

The first stage of the evaluation consisted of the utilization of images of consumable items from the training subset to generate predictions and assess the accuracy of the two models, namely YOLOv5x6 and YOLOv8l. To address the issue of overfitting, both models were trained for a total of 1,000 epochs, with the implementation of early stopping. Early stopping is a regularization technique employed in ML to terminate the training process when the model’s performance on a validation set ceases to improve beyond a predetermined number of epochs. Following the completion of the training procedure, an assessment was conducted to evaluate the efficacy of the model. This evaluation encompassed various metrics, such as precision, recall, and mAP at IoU thresholds of 0.5 (50%) and 0.95 (95%).

Figures 4 and 5 illustrate graphical representations of the metric curves, demonstrating the training progression of the YOLOv5x6 and YOLOv8l models, respectively. The box loss plot displays the loss associated with the accuracy of bounding box coordinates predicted by the model. The objectness loss plot visualizes the change in loss over the training epochs, indicating how effectively the model is learning to distinguish between bounding boxes containing objects and those containing background regions. Both models showing a decreasing objectness loss signify an improvement in their ability to identify objects within the bounding boxes, essential for accurate object detection. The classification loss plot shows the average classification loss value of consumable items. A lower box, objectness, and classification loss signify a higher level of accuracy in the training dataset for consumable items, emphasizing the effectiveness of the model in recognizing and categorizing these items accurately. On the contrary, the validation (val) box loss plot, val objectness loss plot, and val classification loss plot are visual representations of how well models perform on validation data. The precision plot illustrates the accuracy of the model’s positive predictions by showing the ratio of true positive predictions to all positive predictions. In contrast, the recall plot demonstrates the model’s ability to identify all relevant instances by revealing the ratio of true positive predictions to all actual positive instances. Higher precision and recall values signify fewer false positives and the model’s capability to detect a larger proportion of actual positive objects, respectively. The mAP@0.5 plot evaluates performance at a single IoU threshold (0.5) and provides an overall assessment of the model’s performance and how well the predicted bounding boxes overlap with the ground truth boxes. While mAP@0.5:0.95 plot assesses performance across a range of IoU thresholds (0.5–0.95), offering a comprehensive evaluation of the model’s object detection accuracy across various IoU thresholds. In terms of the YOLOv5x6 model, the validation precision score stood at 0.878, with a recall score of 0.973. Additionally, the mAP scores for @0.5IoU and @0.95IoU were 0.935 and 0.905, respectively, as presented in Table 1. On the other hand, the YOLOv8l model exhibited a validation precision score of 0.816, a recall score of 0.971, and mAP scores of 0.925 and 0.915 for @0.5IoU and @0.95IoU, respectively. These outcomes from both models affirm the effectiveness of our approach in accurately monitoring consumables across diverse environments.

Figure 4

Performance metric of YOLOv5x6 feature.

Figure 5

Performance metric of YOLOv8l feature.

Additionally, it is important to note that the YOLOv5x6 model underwent early stopping in its training process at epoch 500, while the YOLOv8l model continued training until epoch 232. In fact, YOLOv5x6 was trained for a longer duration, 500 epochs, before an early stop was initiated. This implies that it continued to make improvements to the validation metric for a more extended period, indicating that it might have been gradually fine-tuning its weights and learning more intricate patterns in the data. In contrast, YOLOv8l ended training significantly earlier at epoch 232, indicating that the model had reached a solution more rapidly. This implies that it reached a point where further training did not significantly improve performance, or it might have encountered challenges like overfitting in the training process that were not encountered by YOLOv5x6.

The model performances not only shed light on the strengths and potential limitations of each model but also underscore the significance of adopting a customized approach to object detection, ensuring accurate monitoring of consumables across diverse environments. This comprehensive evaluation contributes to refining and optimizing object detection models for real-world applications.

4.2 Comparative analysis of system

Furthermore, we investigated the detection capabilities of models with varying sizes. Table 1 displays the detection results of YOLOv5x6 and YOLOv8l on the test set.

When comparing the two models, YOLOv5x6 outperformed YOLOv8l in terms of precision and mAP@0.5, as indicated in Table 1. The precision score of the YOLOv5x6 model was 0.878, while the YOLOv8l model achieved a precision score of 0.816. This suggests that the improved YOLOv5x6 model exhibited superior accuracy in identifying true positives. Both models demonstrated high recall values, signifying their ability to correctly identify the majority of consumables in the images. YOLOv5x6 achieved a recall of 0.973, whereas YOLOv8l achieved a recall of 0.971, indicating that both models successfully detected the majority of consumables in the images.

The mAP is a metric that measures the model’s performance at various IoU thresholds. In this study, both models displayed good mAP@0.5, indicating their high accuracy in object detection. Notably, YOLOv5x6 outperformed YOLOv8l with an mAP@0.5 of 0.935, compared to the YOLOv8l’s score of 0.925. However, when considering the mAP@0.5:0.95 score, YOLOv8l achieved a higher score of 0.915, surpassing YOLOv5x6’s score of 0.905. This implies that YOLOv8l exhibited greater accuracy in detection, particularly at higher IoU thresholds, compared to the improved YOLOv5x6. Moreover, we addressed the substantial difference in training time observed between the two models in our evaluation. Specifically, YOLOv5x6 required 4.580 h for the complete training process, showcasing a significant training time investment. In contrast, the efficiency of YOLOv8l is remarkable, as it completed the entire training process in just 2.010 h. This disparity in training times between the models is a critical consideration in assessing their practicality and resource requirements.

To evaluate the integration of the models in our system, the outcomes of the evaluation of the image training using the YOLOv5x6 and YOLOv8l models are illustrated in Figure 6(a) and (b), respectively. Evidently, the results demonstrate that training with the YOLOv5x6 model produced noticeably superior precision metrics, with values of 0.97 and 0.98. In contrast, the use of the YOLOv8l model for training yielded slightly less precise results, with values between 0.92 and 0.96.

Figure 6

Performance of detection results of the models (a) YOLOv5x6 and (b) YOLOv8l.

Analyzing the YOLOv5x6 and YOLOv8l models reveals distinct performance differences, with YOLOv5x6 demonstrating superior precision and mAP@0.5 despite a longer training duration. These insights offer crucial guidance for selecting models aligned with precision requirements and resource constraints, emphasizing the need for tailored approaches in model selection for effective object detection in diverse environments. The findings contribute to informed decision-making in real-world applications, where achieving the right balance between precision and resource efficiency is pivotal.

5.1 Future directions

To address the identified limitations, future research could focus on developing more adaptive and efficient training processes for object detection models. One important emphasis is on improving training procedures to be more flexible and effective, with a focus on applying the technology in real-world applications. This involves investigating automated techniques for gathering and labeling data, with the goal of optimizing the training process and enhancing the system’s overall reliability. Improving rules and supervision during these processes is essential for ensuring high-quality annotations and enhancing the system’s effectiveness in real-world situations.

In the future, our future scope will be dedicated to the practical application of AI technology in real-world settings, particularly in medical laboratories. Implementing AI systems in real operational settings requires comprehensive training for the medical staff in addition to technological advancements. Therefore, training medical workers is the key importance for optimizing the utilization of AI technologies, enhancing efficiency, and minimizing potential challenges during their implementation. For example, a well-structured training program could include hands-on sessions where medical staff learn to interpret AI-generated understandings, ensuring an integration of AI systems into their daily tasks. This training is crucial for healthcare workers to acquire the required skills and knowledge of AI systems, enabling them to fully connect these technologies and, consequently, enhance the efficiency and workload management in medical laboratories.