MLens: Advancing the Real-Time Detection, Identification, and Counting of Pathogenic Microparasites Through a Web Interface

Abstract

In this study, a diverse collection of images of myxozoans from the genera Henneguya and Myxobolus was created, providing a practical dataset for application in computer vision. Four versions of the YOLOv5 network were tested, achieving an average precision of 97.9%, a recall of 96.7%, and an F1 score of 97%, demonstrating the effectiveness of MLens in the automatic detection of these parasites. These results indicated that machine learning has the potential to make microparasite detection more efficient and less reliant on manual work in parasitology. The beta version of the MLens showed strong performance, and future improvements may include fine-tuning the WebApp hyperparameters, expanding to other myxosporean genera, and refining the model to handle more complex optical microscopy scenarios. This work presented a significant advancement, opening new possibilities for the application of machine learning in parasitology and substantially accelerating parasite detection.

1. Introduction

The class Myxozoa comprises over 70 genera and approximately 2600 species, all of which are obligate endoparasites [1,2,3]. Many of these parasites cause severe diseases known as myxosporidioses. Lom and Arthur [4] described morphological and morphometric variables as indispensable factors in defining myxospores and taxonomic groupings.

Among myxozoans, the genera Henneguya [4] and Myxobolus play significant roles as parasites of ichthyofauna [5]. These polyphyletic genera encompass various members distributed worldwide, are found in both freshwater and marine ichthyofauna, and have also been reported in reptiles and amphibians [1,6]. Their geographic distribution is influenced by environmental factors such as water temperature, water quality, and host availability [7,8,9].

Neural networks aspire to emulate the complexity of biological neuronal functions [10]. Analogous to cellular processes, artificial neurons excel in their ability to receive, process, and transmit information through mathematical operations, thereby enabling them to identify patterns, perform classifications, and predict characteristics [10,11]. This biology–mathematics symbiosis-inspired process opens promising horizons for innovation in the field of AI, where human and artificial minds intertwine in the pursuit of innovative and bold solutions [12].

The application of AI finds fertile ground in parasitology, playing a crucial role in addressing persistent challenges, and convolutional neural networks are tools used in computer vision (CV) and are now essential in veterinary medicine and industry [13]. They are intrinsically associated with CV, which, when combined with image analysis techniques, results in comprehensive specialized systems that offer solutions for detecting and classifying parasitic objects with minimal human intervention and high accuracy [14]. This, in turn, eliminates the errors often associated with human analysis stemming from tasks that are often fatiguing, monotonous, costly, or simply lack expertise [15,16,17].

Moreover, the acquisition of microscopy images establishes a solid foundation for feeding CV systems, revealing new diagnostic and research perspectives in parasitology [18,19].

In this study, a beta version of a WebApp called MLens 1.0.0 beta was developed to perform the automated detection and classification of the genera Myxobolus and Henneguya in light microscopy-acquired images via the YOLOv5 neural network, aiming to increase the efficiency and accuracy of parasitological diagnosis.

Background

Advances in object detection using algorithms such as YOLOv5 emerged as an effective approach for complex identification tasks in optical microscopy images [20]. Networks like these offer fast and accurate solutions, minimizing human intervention and maximizing accuracy even in challenging environments [21]. Previous studies demonstrated the effectiveness of these tools in biology and medicine, including the detection of cellular structures and microparasites [14,22].

As highlighted by Sichman et al. [23], humanity’s relentless pursuit of new heights of progress is undeniably driven by continuous advances in the broad domain of science, as evidenced in the case of artificial intelligence (AI), which played a pivotal role in enabling the development of specialized computational technologies. Therefore, this phenomenon sparked substantial and enduring interest, with its impact continuing to resonate notably in an ever-evolving society [21,24]. However, their application to the diagnosis of myxozoans is still emerging, with gaps related to model generalization in heterogeneous scenarios and the morphological variability of parasites [25,26].

This article was structured into five main sections. In Section 2, we detailed the materials and methods used, including the data collection, preparation, and training process with the YOLOv5 algorithm. Section 3 presented the results obtained and an in-depth discussion of this study’s implications and limitations, highlighting the main findings [27,28]. Section 4 concluded this article and explored future perspectives, At the end of this article, the references that underpin this work [29,30] are listed. This organization aimed to provide a clear and objective view of the MLens system’s contribution to modern parasitology.

2. Materials and Methods

The development of the MLens 1.0.0 beta was carried out in Belém, PA, Brazil or processing light microscopy images, aimed at detecting two genera of myxozoans via the You Only Look Once (YOLOv5) algorithm, was performed in three distinct phases: dataset collection, project dataset training, and implementation. Notably, all stages were conducted at the Laboratory of Morphomolecular Integration and Technologies (LIMT) of the Federal Rural University of the Amazon.

2.1. Cnidaria: Myxozoa

Myxozoa (Cnidaria: Endocnidoxoa) represents one-fifth of the known cnidarian species [25,31] worldwide, spanning freshwater and marine environments [31,32]. Many infections are asymptomatic and are considered harmless. However, some species have high pathogenic potential, causing disease and significant mortality in parasitized organisms. This can lead to substantial economic losses in the fishing and aquaculture sector. Additionally, immunocompromised individuals who consume raw fish protein from infected fish may experience harm caused by these parasites, although there is no evidence that humans are hosts in any stage of the myxozoan life cycle. However, myxozoan spores were detected in human feces [33].

Myxozoans exhibit a complex life cycle that involves alternating between invertebrate and vertebrate hosts. The latter serve as intermediate hosts, whereas invertebrates act as definitive hosts, particularly those from the classes Oligochaeta and Polychaeta, which belong to the phylum Annelida [7]. Transmission between hosts occurs through two morphologically distinct forms: the myxospore, which develops in vertebrates, and the actinospore, which is found in invertebrates. Typically, this transmission occurs via spores that disperse in water, as shown in the work of Fiala et al. [2] and Figure 1.

Okamura et al. [25] reported that myxozoans exhibit pronounced structural simplification, along with miniaturization, as a result of their parasitic niche. Okamura et al. [7] reported that these parasites obligatorily undergo complex life cycles, requiring both invertebrate and vertebrate hosts. This extensive group of endoparasites has complex, obligatory microscopic, and heterogeneous life cycles that are dependent on both invertebrate and vertebrate hosts [25,34]. These parasites have the potential to cause significant environmental and economic damage. Although they are commonly found in fish, they can also infect reptiles, amphibians, waterfowl, and even small mammals, as shown in the work of Eiras et al. [26].

2.2. Data Acquisition

The image collection system used in this study belongs to the LIMMT. The team imaged all parasites of the genera Henneguya and Myxobolus found during necropsies between 2022 and 2024 via smartphones, as exemplified in the pipeline (Figure 2). In total, 125 images with dimensions of 2160 × 3840 pixels were collected and subdivided into 12 smaller images with dimensions of 640 pixels. Therefore, the sample comprised 12,000 images divided into two genera, Myxobolus and Henneguya, covering various species of Amazonian fish and different tissues, such as gills, liver, and skeletal muscle. The images were uniformly sized at 640 pixels each and were in JPEG format.

2.3. Data Preprocessing

To enhance the generalizability of the model for identifying microparasites in microscopy images, several image processing techniques were employed, as shown in Figure 3. As discussed by Zhao et al. [20], rotations were applied at angles of 90°, 180°, and 270°. Additionally, horizontal and vertical flips were employed on the original images and at 90° clockwise, as demonstrated in the works of Bydder et al. [35], Zendehdel et al. [22], and Ghose et al. [36], thus avoiding redundancies. These transformations provided increased diversity in the dataset because, after subdividing the original images, each subunit resulted in eight new unique images. This helped the model recognize patterns in multiple orientations and reflections, thereby improving its robustness and accuracy in detecting these genera in various scenarios.

2.4. Labeling of Myxozoans

The Roboflow platform for CV was used to demarcate the regions of interest, that is, to label the parasites in the form of bounding boxes. As emphasized by Noever and Noever [16], the essence of this task was the accuracy of labeling the objects of interest, which, in turn, depended on human performance. As illustrated in Figure 4, the quality and consistency of these labels were crucial because they served as the foundation for the learning process of the model for identifying and locating objects in new images.

This requirement was met by carefully creating bounding boxes, precisely outlining the location of the myxozoans, and assigning each genus a specific label, as shown in

Table 1. Labeling scheme in text file—X and Y are the center coordinates of the target, and W and H are the width and height of the target, respectively.

Genera	X	Y	W	H
1	0.72500000	0.69609375	0.11718750	0.79609375
1	0.83125000	0.78437500	0.07343750	0.48437500
0	0.44140625	0.64765625	0.10625000	0.74765625

. This enabled the generation of label files in the “.txt” format linked to the source image. These files included the category (in this case, the genus) and bounding box coordinate information (X: x-coordinate center of the bounding box; Y: y-coordinate center of the bounding box; W: width of the bounding box; and H: height of the bounding box), as illustrated in Table 1.

The model was trained after appropriately labeling the data and organizing the training and validation sets. MLens was subsequently implemented in a real-world context by performing myxozoan detection on the basis of the meticulously provided information in this process.

Table 2. Information on each of the genera labeled for YOLO: Henneguya and Myxobolus.

Category Information Genera Labeled for YOLOv5
Henneguya	0		Myxobolus	1

shows the categories listed in Table 1.

Furthermore, this study aimed to obtain reliable evidence that the labeled genera Henneguya and Myxobolus presented a numerical balance. The analysis included an exploration of the characteristics of the labels used and the associated bounding boxes (Figure 5—statistical analysis of labels: gender histogram and density map of bounding box sizes). These genera exhibited a quantitative balance. However, the number of parasites in question tended to be very small compared with the image scale, which could complicate detection.

2.5. Quantitative Characteristics of the Dataset

The images were randomly distributed for training and validation at a ratio of 80:20. However, when making such divisions, obtaining a balance between the number of samples of each genus in the training and test sets was important to ensure maximum consistency in data distribution.

2.6. Experimental Configuration

To ensure the rigor of this study, Google Colaboratory was used for the training and validation phases of the model. This free cloud server was designed to foster research and education in machine learning and includes integrated libraries, such as TensorFlow, Matplotlib, Keras, and OpenCV, eliminating the need for complex local installations, as demonstrated by Ray et al. [37]. Additionally, Colab facilitates synchronization with Google Drive, allowing the storage of large data volumes without consuming disk space. It also promotes mobility, collaboration, and project sharing, requiring only an email account, making it convenient for community collaboration.

This platform provides hardware adapted for machine learning, including an Intel Xeon CPU; an NVIDIA Tesla^® L4 GPU with 24 Gigabytes (GB) of GDDR6 memory and 7680 CUDA cores (NVIDIA, Santa Clara, CA, USA); and 78 GB of storage and 51 GB of RAM, which is supported by CUDA v12.2. The deep learning frames PyTorch v2.2.1 and Python v3.10.12 were used for development. Finally, an interactive dashboard was created via Streamlit v1.19.0.

2.7. YOLOv5

YOLOv5 has become an essential algorithm in contemporary object detection, balancing speed and execution accuracy, as demonstrated in the works of Liu et al. [38] and Li et al. [21]. Object detection is a fundamental digital field. It involves identifying the objects present in an image or frame and determining their exact locations. This type of activity is more complex than object classification, which only recognizes the presence of objects without indicating their positions in an image, as shown by Xu et al. [39]. Moreover, objects cannot be classified multiple times, increasing the complexity of the detection process.

The range of YOLOv5 is noteworthy. Compared with previous versions, YOLOv3 and YOLOv4, YOLOv5 boasts higher accuracy and excels in terms of lightness, speed of object detection, and inference speed. These characteristics are crucial for embedded applications that require real-time detection, as demonstrated by Xu et al. [39], Zhang et al. [40], and Badgujar et al. [41]. These findings demonstrated that this algorithm can substantially improve the detection and identification of myxozoans via light microscopy.

The YOLOv5 architecture includes various variants; however, in this study, only the following four variants were used: YOLOv5n, YOLOv5s, YOLOv5m, and YOLOv5l. As discussed by Elshahawy et al. [27], the recommended steps for machine training were followed and documented in a notebook hosted on Google Colab. However, its ability and adaptability to other applications using different datasets are relatively simple, flexible, and open.

2.8. Evaluation Metrics

In machine learning, the interpretation of appropriate metrics ensures the accurate assessment of a model. However, these metrics can vary from situation to situation, causing some difficulty in comparison and potentially leading to erroneous conclusions. In this study, the precision (P), recall, F1_Score, and mAP were used as evaluation metrics [42,43].

According to Shen et al. [28], the mAP metric is the average precision calculated across all genera for all categories (N). Notably, an mAP of 0.5 is an intersection over the union (IoU) threshold of 0.5. Furthermore, mAP0.5:0.95 represents the average across IoU thresholds from 0.5 to 0.95, with an interval of 0.05. The formulas for precision, recall, mAP, and F1_Score are presented in the following section.

3. Results and Discussion

Standard hyperparameters were used: an initial learning rate (lr0) of 0.01 and a final learning rate (lrf) of 0.01. Furthermore, an SGD with a momentum of 0.937 and weight_decay of 0.0005 was used to control the regularization and avoid overfitting. A batch size of 16 was selected to balance the training speed and model accuracy [44], with training lasting for 300 epochs to allow the models to clearly absorb patterns in the data, and the image dimensions were set to 640 pixels. In the first three epochs, a warmup_momentum of 0.8 and warmup_bias_lr of 0.1 were applied.

The emergence of various object detection models, such as the YOLO series, revolutionized the field of CV. Each variant of these networks has unique characteristics in terms of GPU memory requirements [29,45,46,47]. The results were derived from the detection training of the Henneguya and Myxobolus genera via the YOLOv5n, YOLOv5s, YOLOv5m, and YOLOv5l variants configured with the parameters listed in Table 1, as exemplified by Huang et al. [48] and Du [30].

In particular, the YOLOv5 model has drawn significant attention because of its high performance and efficiency. In Figure 6, the YOLOv5l model requires 11 GB of GPU memory to stabilize its training process, whereas the YOLOv5m model requires 7.1 GB. Interestingly, the YOLOv5s variant requires only 4 GB of GPU memory, and the most resource-efficient model, YOLOv5n, uses only 2.4 GB of GPU memory until the completion of training [47,49].

The most plausible explanation for the difference in GPU memory requirements can be attributed to the inherent architectural characteristics of each variant. As discussed by Hasan et al. [47], the YOLOv5n model, which is the smallest in the series, is more suitable for deployment on mobile hardware platforms, such as those embedded in vehicles, owing to its compact size and low memory demand.

Nevertheless, the analysis of the training time and model performance is a common consideration in machine learning. Larger models, such as YOLOv5l, are often more accurate in detection but are associated with robust computational resources and longer training times [50]. Fu et al. [51] highlighted that smaller variants such as YOLOv5n are more efficient and suitable for implementation on resource-constrained platforms while accepting some precision trade-offs.

The variants differed in their training completion times, as shown in Figure 6. Notably, YOLOv5n, the shortest, took 400.46 min, followed by YOLOv5s (508.85 min) and YOLOv5m (870.15 min), and YOLOv5l, which was the most complex, required 1336.78 min. Additionally, the model sizes, presented in Figure 6, corresponded to their complexities: YOLOv5n resulted in a 3.9 MB model, YOLOv5s in a 13.8 MB model, YOLOv5m in a 40.3 MB model, and YOLOv5l in a 92.9 MB model. These differences clearly indicate that model complexity influences training time and hardware demand, as mentioned by Tong and Yiquan [50] and Cai et al. [52].

The training performance curves of the models are shown in Figure 7a–d. Ma et al. [53] reported that such metrics are evaluation criteria for models that allow for the understanding of their advantages and disadvantages and comparison of different models to determine the most suitable model for real-world situations. Without such metrics, verifying the progress or superiority of one model over others would be impossible [54,55].

According to Figure 7a,c, the precision and recall rates of all the models started to increase significantly from epoch 2. Initially, at epoch 0, the precision and recall varied according to the complexity of the model: YOLOv5l started with the highest precision at 59.6% and recall at 68.7%; YOLOv5m had the highest initial precision, reaching 69.6% and a recall of 68.1%, whereas YOLOv5n had the lowest values (41.5% and 58.6%, respectively). All the models achieved accelerated gains in precision and recall for up to approximately epoch 10.

The mAP scores, shown in Figure 7b, demonstrated a similar trend. The YOLOv5m and YOLOv5l models begin with higher mAP_50 scores (66.8% and 71.8%, respectively); however, the YOLOv5l model outperforms YOLOv5m (93.4% and 92.6%, respectively) only at approximately epoch 18. However, until epoch 10, all the variants rapidly evolve in terms of mAP_50.

When the mAP50:95 metric was evaluated via Figure 7d, an interesting dynamic was observed between the YOLOv5l and YOLOv5m variants. The training started with YOLOv5m at an mAP50:95 of 38.7%, followed by YOLOv5l at 36.5%. Both methods achieved similar performances until epoch 9, when they reached mAP50:95 values of 64.5% and 66.3%, respectively. However, as the training progressed, YOLOv5l showed consistent improvement, achieving a remarkable score of 89.2% at the end of the training. However, YOLOv5m maintained considerable performance, scoring 86%. In contrast, YOLOv5s reached 80.6%, and YOLOv5n, despite improving over the epochs, lagged behind with only 73.5% mAP50:95.

Figure 8a–c illustrates the training losses, highlighting the effectiveness of YOLOv5 variants over 300 epochs in detecting myxozoans. Initially, all the models experienced abrupt losses, which decreased rapidly during the first 50 epochs and then gradually stabilized. Among them, YOLOv5n, being the least complex, showed greater losses and resulted in greater losses than did other more complex models, such as YOLOv5s, YOLOv5m, and YOLOv5l, which showed a better fit to the training data. This difference underscores the superior ability of more robust models to capture data patterns.

However, Figure 9a–c indicates losses during validation, which were not discrepant and followed a trend similar to that during training, with the losses rapidly decreasing in the first 50 epochs. The YOLOv5n model again presented the highest final losses, whereas the more complex models achieved better validation results.

These results reflected the superior generalization capacity of the more complex models, which is essential for effective performance on new data, as shown in

Table 3. YOLOv5 model performance by each genus.

Model	Genus	Images	Parasites	Precision	Recall	mAP:50	mAP:50–95
YOLOv5l	All	2000	12,386	0.974	0.967	0.980	0.892
	Henneguya	2000	5396	0.994	0.970	0.993	0.905
	Myxobolus	2000	6990	0.954	0.965	0.966	0.879
YOLOv5m	All	2000	12,386	0.972	0.957	0.976	0.860
	Henneguya	2000	5396	0.995	0.951	0.990	0.876
	Myxobolus	2000	6990	0.949	0.962	0.962	0.844
YOLOv5n	All	2000	12,386	0.928	0.895	0.943	0.727
	Henneguya	2000	5396	0.976	0.852	0.947	0.74
	Myxobolus	2000	6990	0.880	0.938	0.939	0.713
YOLOv5s	All	2000	12,386	0.948	0.935	0.964	0.954
	Henneguya	2000	5396	0.985	0.915	0.974	0.815
	Myxobolus	2000	6990	0.911	0.954	0.954	0.780

, with precision, recall, mAP:50, and mAP:50–95 for the validation set.

Figure 10a shows precision and recall, while Figure 10b highlights precision and confidence, the performance of the models in distinguishing the genera Henneguya, Myxobolus, and background. Overall, the YOLOv5l model achieved a robust accuracy of 99.3% for Henneguya, indicating that almost all of its predictions were correct, even with high recall. For Myxobolus, a precision of 96.6% demonstrated high performance, though slightly lower than that of Henneguya. The average precision for all classes (mAP: 0.5) of 97.9% highlights the consistency of the model. In the recall–confidence curve, even with high confidence, the model maintains recall, indicating that its predictions are reliable.

The confusion matrices shown in Figure 11a–d for the YOLOv5n, YOLOv5s, YOLOv5l, and YOLOv5m variants provide an analysis of the performance for the detection of Henneguya, Myxobolus, and background. Furthermore, YOLOv5n correctly identified 4.89k (TP) instances of Henneguya, with 396 false positives (FPs), and 6.70k TPs for Myxobolus, with 954 FPs. Although it performs well in detection, the number of FPs can still be reduced. Additionally, YOLOv5s maintained balance, identifying 5.13k TPs for Henneguya with 274 FPs and 6.78k TPs for Myxobolus with 648 FPs. A notable reduction in false positives was observed compared with YOLOv5n.

Finally, YOLOv5m stood out with 5.23k TP for Henneguya and only 153 FPs and 6.82k TP for Myxobolus with 534 FPs, showing a good balance between precision and discrimination capability. However, YOLOv5l demonstrated significant superiority, with 5.29k TP for Henneguya and only 130 FPs and 6.83k TP for Myxobolus, with 433 FPs, indicating its higher precision.

Figure 12 provides a qualitative analysis of the detection results for the genera Myxobolus and Henneguya for some images not used in the proposed dataset. As depicted in the images, myxozoans of various shapes and positions were observed, though belonging to the genera Myxobolus and Henneguya, which were still detected with high precision. This demonstrates that the MLens can perform robust detections.

4. Conclusions

In the present study, a diverse collection of myxozoan images was obtained from the genera Henneguya and Myxobolus. This collection aimed to cover various situations and scenarios in light microscopy, providing a practical dataset for use in CV. Simultaneously, different versions of the YOLOv5 network were implemented to detect tar-infected parasites, which significantly improved the myxozoan detection accuracy.

Four networks were tested to automate the detection and differentiation of myxozoans and obtained satisfactory results, achieving an average precision of 97.9% for all genera, with a recall of 96.7% and an F1_Score of 97%. These results demonstrated the potential applicability of the MLens in real-world scenarios, indicating that machine learning in parasitology offers considerable benefits, making microparasite detection processes more efficient and less labor intensive.

From a forward-looking perspective, further advancements in web and mobile applications for the detection of myxozoans and parasites aimed at increasing the number of identifiable genera with continuous improvements in precision, accuracy, and real-time performance are expected. This is an important field for future research, including topics such as enhancing the quality and diversity of datasets and addressing the challenges in detecting multiple small targets in complex scenarios.

The MLens showed excellent results as a beta version, and further medium-term training, fine tuning of hyperparameters, and expansion of the collection are needed. This is the first software developed for the detection of myxozoans, which will hopefully serve as a gateway to the field of parasitology associated with AI. Building upon the success of this study, future work aims to develop more specific versions of the application, explore novel applications of the MLens, and refine its use for broader and more specialized parasitological diagnoses, fostering innovation in the intersection of artificial intelligence and parasitology.