Artificial Intelligence with Convolutional Neural Networks for Microfilariae Detection in the Brazilian Amazon

Abstract

The Amazon region faces persistent structural limitations for the diagnosis of filarial diseases. This study aimed to develop and evaluate an artificial intelligence model based on convolutional neural networks to classify microscopic images according to the presence or absence of microfilariae. This was a technological, quantitative, and applied study in which blood samples from 43 dogs were collected in rural areas of Manaus, prepared on stained slides, and digitized using a webcam coupled to a microscope, generating 500 original images. The images were preprocessed, organized into binary classes, and subjected to data augmentation in the training set, resulting in approximately 1,000 instances. Ground truth was established through expert morphological assessment and molecular confirmation by laser microdissection and polymerase chain reaction (PCR). The EfficientNetV2-B0 model, trained using a patch-based approach, achieved an accuracy of 93.6%, precision of 91.8%, sensitivity of 92.4%, and an F1-score of 92.1%. The average analysis time per slide was 104 seconds using artificial intelligence, compared with 2,065 seconds for human reading, demonstrating a substantial gain in efficiency and highlighting the potential application of this approach in parasitological screening and epidemiological surveillance.

Key words:
Artificial intelligence; Filariasis; Microfilariae; Amazon; Health surveillance

Resumo

A Amazônia apresenta limitações estruturais persistentes para o diagnóstico das filarioses. Este estudo teve como objetivo desenvolver e avaliar um modelo de inteligência artificial baseado em redes neurais convolucionais para classificar imagens microscópicas quanto à presença ou ausência de microfilárias. Trata-se de um estudo tecnológico, quantitativo e aplicado, no qual amostras sanguíneas de 43 cães foram coletadas na zona rural de Manaus, preparadas em lâminas coradas e digitalizadas por webcam acoplada ao microscópio, gerando 500 imagens originais. As imagens foram pré-processadas, organizadas em classes binárias e submetidas a aumento de dados no conjunto de treinamento, resultando em aproximadamente 1.000 instâncias. O ground truth foi definido por avaliação morfológica especializada e confirmação molecular por microdissecação a laser e PCR. O modelo EfficientNetV2-B0, treinado em abordagem patch-based, alcançou acurácia de 93,6%, precisão de 91,8%, sensibilidade de 92,4% e F1-score de 92,1%. O tempo médio de análise por lâmina foi de 104 segundos com a IA, frente a 2.065 segundos na leitura humana, evidenciando ganho expressivo de eficiência e potencial de aplicação na triagem parasitológica e na vigilância epidemiológica.

Palavras-chave:
Inteligência artificial; Filariose; Microfilárias; Amazônia; Vigilância em saúde

Resumen

La Amazonía presenta limitaciones estructurales para el diagnóstico de las filariasis. Este estudio tuvo como objetivo desarrollar y evaluar un modelo de inteligencia artificial basado en redes neuronales convolucionales para clasificar imágenes microscópicas según la presencia o ausencia de microfilarias. Se trata de un estudio tecnológico, cuantitativo y aplicado, en el cual se recolectaron muestras de sangre de 43 perros en la zona rural de Manaus, preparadas en láminas teñidas y digitalizadas mediante una cámara web acoplada al microscopio, generando 500 imágenes originales. Las imágenes fueron preprocesadas, organizadas en clases binarias y sometidas a técnicas de aumento de datos en el conjunto de entrenamiento, resultando en aproximadamente 1.000 instancias. El ground truth se definió mediante evaluación morfológica especializada y confirmación molecular por microdisección láser y PCR. El modelo EfficientNetV2-B0, entrenado mediante un enfoque patch-based, alcanzó una exactitud del 93,6%, precisión del 91,8%, sensibilidad del 92,4% y un F1-score del 92,1%. El tiempo promedio de análisis por lámina fue de 104 segundos con la IA, frente a 2.065 segundos en la lectura humana, lo que evidencia una ganancia de eficiencia y un alto potencial de aplicación en la tamización parasitológica y la vigilancia epidemiológica.

Palabras clave:
Inteligencia artificial; Filariasis; Microfilarias; Amazonía; Vigilancia en salud

Introduction

The Brazilian Amazon is recognized for its vast territory and its strategic role in global climate balance. It faces persistent structural barriers to healthcare access¹. These challenges are conditioned by geographic, logistical, and socioeconomic factors and directly impact the surveillance, diagnosis, and control of neglected tropical diseases, particularly in territories characterized by population dispersion, dependence on river transport, limited connectivity, and scarce laboratory infrastructure^2-4.

Among the Neglected tropical diseases (NTDs), filarial diseases remain a significant public health problem in the Amazon region, both due to their clinical impact and the underreporting associated with local diagnostic limitations. The required microscopy for identifying microfilariae depends on trained personnel and proper equipment-resources often lacking in low-complexity health settings. This reality leads to diagnostic delays, sustained transmission, and weak epidemiological surveillance^5,6.

While the issue of underdiagnosis persists, this problem has not been historically ignored. The first documented case of Dirofilaria immitis in Manaus dates to 1922, identified during necropsies on stray dogs at the Municipal Incinerator facility⁷. This remained an isolated finding until a recent hematological survey discovered microfilariae in domestic dogs, renewing epidemiological interest and debate⁸.

Dirofilariasis is caused by Dirofilaria immitis and is commonly known as heartworm disease. It is a mosquito-borne zoonosis transmitted by mosquitoes from the genera Aedes, Anopheles, and Culex. Dogs serve as the primary hosts, but incidental human infections may occur. These can result in pulmonary nodules that mimic neoplasms-often described as coin-like lesions^9-11. Such cases may trigger unnecessary oncologic investigations and increase healthcare costs within Brazil’s Unified Health System (SUS). The risk of misdiagnosis is compounded by the limitations of local laboratories. Parasite identification requires specialized training, appropriate microscopy, and often molecular confirmation¹².

Recent evidence further supports the persistence of transmission in the region. Dirofilaria immitis remains endemic in rural areas of Manaus. Prevalence is higher than in urban and peri-urban settings. Studies from 2017 to 2021 documented a consistent spatial gradient. Urban areas showed low prevalence (0.35%), peri-urban areas were higher (1.22%), and rural areas were substantially higher (15.44%). This pattern suggests more intense transmission in rural territories. These areas, thus, may play a central role in maintaining the infection cycle in the Amazon region⁸.

Against this backdrop, expeditions to the Tupé Sustainable Development Reserve (RDS) were conducted between 2022 and 2025, during which systematic blood collection and the recording of microfilariae in microscopic preparations was carried out. The material collected during these expeditions gave rise to the image bank used in this study, built under real field conditions compatible with the routine of health units located in remote areas.

Even when samples are available, however, microscopic identification poses additional challenges. Distinguishing microfilariae from artifacts such as fibrin strands, debris, and cellular remnants is particularly difficult because these structures can adopt elongated shapes similar to those of the parasite (Figure 1). This visual overlap is more common at lower magnifications or when lower-resolution equipment is used, making interpretation highly dependent on the examiner’s experience, lighting conditions, and image contrast. Under such circumstances, interpretive variability increases, contributing to false-positive and false-negative findings and underscoring the need for complementary diagnostic approaches.

Figure 1
Blood smear stained with Giemsa, showing thread-like structures compatible with microfilariae (white rectangles) and morphologically similar artifacts (black rectangles). Image obtained by low-magnification optical microscopy.

Although diagnostic accuracy tends to improve at higher magnifications, such as 40× or 63×, where contour, nuclear arrangement, and cytoplasmic texture become visible, lower magnifications place greater weight on the examiner’s experience and on optical conditions. These limitations are even more critical in settings where adequate equipment and specialized personnel are scarce, a reality common in many areas of the Amazon¹³.

Amid these constraints, recent advances in artificial intelligence have drawn increasing interest in automated biomedical structure recognition. They offer potential to improve standardization, accelerate diagnostics, and reduce dependence on highly trained specialists^14-17. In filarial infections, however, major challenges remain. Low parasite density, morphological variation, and the need for robust annotated image sets make model training and generalization difficult^18,19.

A recent review identified 283 studies that applied artificial intelligence to peripheral blood smear analysis. Most focused on malaria, leukemias, leukocytes, mixed cells, erythrocytes, and myelodysplastic syndromes. None addressed the detection of microfilariae using artificial intelligence. This highlights an important gap in the international literature²⁰. Ordinance GM/MS No. 3,691/2024 and the SUS Digital Health Strategy reinforce the need for scalable, low-cost solutions. These technologies can be used where infrastructure is limited and connectivity is unstable, conditions typical in the Amazon²¹.

Based on these unmet institutional and scientific needs, this study proposes the development and evaluation of an artificial intelligence model using convolutional neural networks for detecting microfilariae in microscopic images with a patch-based approach. By aligning with the operational realities of low- and medium-complexity health services, this solution is designed to strengthen filarial surveillance in the Amazon, supporting-rather than replacing-expert assessment.

Methods

Study type and design

This study was designed to develop and evaluate an artificial intelligence model for automated detection of microfilariae in optical microscopy images. It is an applied technological development study with a quantitative approach. The study integrates biological, laboratory, molecular, and computational stages. It covers field sample collection to model validation in a simulated routine laboratory setting.

Samples were collected as part of a zoonotic surveillance project and after informed consent was obtained from animals’ guardians. Samples were labelled with alphanumeric codes and were unlinked to personally identifiable data. This approach ensured anonymization and prevented individual traceability.

Data

Field activities were conducted in the Tupé Sustainable Development Reserve (RDS), a riverine area in the rural part of Manaus, the capital of Amazonas. Access is primarily by river and, secondarily, by land, depending on seasonal hydrological variation. Blood samples were collected from 43 dogs. They were stored in EDTA tubes and kept refrigerated until laboratory processing.

For microscopic preparation, blood aliquots were placed on slides with a PEN membrane, stained, and examined by bright-field optical microscopy. Initial parasitological screening used classic morphological criteria such as length, thickness, nuclear pattern, and overall appearance of the filiform body. There was considerable diagnostic ambiguity because microfilariae resemble non-parasitic structures like fibers, fibrin, cellular debris, and staining artifacts, which reinforces the need for subsequent molecular confirmation.

The image bank was built from 50 blood smear slides made from the collected samples and processed under routine laboratory conditions. For comparative analysis between human reading and the automated system, nine slides were selected by simple random sampling from the eligible pool. This pool was predefined by minimum optical quality criteria and representation of different parasite burdens. The full set of slides supported the dataset for training, validation, and testing of the artificial intelligence model.

Microscopic images were digitized using a bright-field optical microscope coupled to a Logitech C920 webcam at a resolution of 1920×1080 pixels. This yielded 500 original microscopic images obtained under real screening conditions. The initial dataset was organized into two main classes: presence and absence of microfilariae. The classes included negative fields and artifactual structures, reflecting the morphological heterogeneity seen in routine diagnosis. Images were submitted to computational preprocessing, which included normalization to the [0.1] range and resizing to 224×224 pixels to match the architecture’s input format.

Data augmentation techniques were applied exclusively to the training set, including rotations, flips, brightness and contrast adjustments, and zoom operations, generating synthetic variants of the original images. This procedure produced an effective training set of approximately 1,000 images, increasing the morphological variability presented to the model and reducing the risk of overfitting without introducing new biological data. The dataset was then partitioned into training (70%), validation (15%), and test (15%) sets, aligned with established good practices in biomedical artificial intelligence and ensuring a robust, unbiased assessment of model performance^22,23.

Reference test (ground truth)

Ground truth was established through a highly reliable multimodal protocol that combined expert morphological assessment with molecular confirmation.

First, stained slides were independently examined by researchers with expertise in parasitology, who identified regions suggestive of microfilariae based on classic morphological criteria. The selected structures then underwent laser microdissection (LMD), DNA extraction using a low-volume protocol, and PCR amplification with the ITS-1 marker, allowing unequivocal molecular confirmation.

Only regions and images whose presence or absence of microfilariae was confirmed through this integrated workflow were included in the final dataset; ambiguous regions or regions lacking molecular confirmation were excluded. In the computational workflow, ground truth was applied at the image-region level using a patch-based strategy, in which each patch received a binary label (presence/absence of microfilariae). This approach enabled a granular and realistic evaluation that reflected the heterogeneity of microscopic fields observed in laboratory practice (Figure 2).

Figure 2
Steps in the laboratory workflow. From left to right: (A) slide with stained smear; (B) laser microdissection; (C) DNA extraction.

Model architecture and training

The model was developed using the EfficientNetV2-B0 architecture with ImageNet-pretrained weights and a transfer learning strategy, and it was configured as a binary classifier to distinguish patches with microfilariae from those without them. The aim was to maximize performance while preserving computational feasibility in low-resource settings.

Training ran for 100 epochs with a batch size of 16, binary cross-entropy loss, and the Adam optimizer (learning rate 1×10^-⁴). Dropout and early stopping were used to mitigate overfitting. The computational environment included Ubuntu 22.04, Python 3.10, TensorFlow/Keras, and Anaconda, with execution on a notebook equipped with an NVIDIA RTX 3050 GPU accelerated via CUDA.

Although k-fold cross-validation is widely used in machine learning, this study adopted a fixed split into training, validation, and test sets. This decision was made because of the risk of information leakage, as multiple patches may be extracted from the same slide. If patches derived from the same slide are present in both training and test sets, performance metrics may be artificially inflated, compromising the validity of the results. In addition, the fixed split was chosen to simulate a realistic operational setting in which the model is trained once and then applied to previously unseen data.

To ensure reproducibility, random seeds were controlled across the main modules, and hardware and software configurations were kept constant throughout the experimental process.

Model inference

At the inference stage, a patch-based approach was adopted in which each frame captured by the webcam attached to the microscope was automatically divided into multiple patches, allowing local analysis of specific regions of the microscopic field. This strategy increases sensitivity to filiform structures and reduces the influence of variation in focus, illumination, and artifacts, making it suitable for elongated and visually subtle organisms such as microfilariae.

Real-time inference was performed exclusively on data not used during training and validation, including slides from the test set and additional samples acquired after the training phase, thereby simulating real-world use. The model was integrated into an inference system with OpenCV for continuous image capture. Each frame was preprocessed, segmented into patches, and classified with an estimated probability displayed as an overlay on the microscopic field. The average processing rate ranged from 5 to 8 frames per second (Figure 3).

Figure 3
AI model in a notebook with an attached webcam, capturing microscopic images displayed on the monitor.

Statistical analysis

Performance was assessed using confusion-matrix metrics, including accuracy, precision, sensitivity (recall), specificity, and F1-score. The ROC curve and the area under the curve (AUC) were used to evaluate overall discriminatory performance. Uncertainty around the metrics was estimated through 95% confidence intervals calculated by bootstrap resampling with 1,000 iterations on the test set, with metrics recalculated at each resampling step.

Results

Comparative performance in slide analysis

As shown in Table 1, the mean accuracy among the human researchers was 87.39%, whereas the artificial intelligence system reached 93.59%, with less than two percentage points of variation across the different slides analyzed. Operationally, mean analysis time per slide was substantially longer for human observers (2,191 s, approximately 36 min) than for the AI system (103 s, about 1 min 43 s), representing an efficiency gain of roughly 21-fold.

Table 1
Performance of Human Researchers and AI in Microfilaria Analysis.

Slide	Amt. MF	T1 (s)	T2 (s)	T3 (s)	T4 (s)	Mean R (s)	T. AI (s)	Ac. H (%)	Ac. AI (%)
01	06	927.0	954.0	1142.0	1068.0	1022.8	117.0	87.7	92.9
02	07	1320.0	1247.0	1336.0	1155.0	1264.5	107.0	87.1	94.0
03	04	797.0	752.0	873.0	782.0	801.0	107.0	88.2	94.4
04	17	3702.0	3737.0	3605.0	3678.0	3680.5	101.0	88.6	93.0
05	36	7589.0	7284.0	7704.0	7699.0	7569.0	117.0	88.3	93.6
06	17	3694.0	3579.0	3716.0	3737.0	3681.5	88.0	88.0	93.1
07	04	783.0	804.0	712.0	686.0	746.3	104.0	86.9	93.7
08	03	540.0	551.0	498.0	542.0	532.8	98.0	85.2	93.8
09	02	405.0	421.0	435.0	430.0	422.8	87.0	86.8	94.2

• Notes: Slide: Identification number of the analyzed slide. Amt MF: Amount of microfilariae observed on the slide. T1 to T4 (s): Analysis time in seconds for researchers 1 to 4. Mean P. (s): Mean time among the four human researchers. AI T. (s): Analysis time of the artificial intelligence system. Ac. H (%): Mean accuracy of the human researchers. Ac. AI (%): Accuracy of the artificial intelligence system.
• Source: Authors (2025).

The analysis of temporal variability further highlights this difference. For AI, processing time ranged from 87 to 117 s, with a standard deviation of ±10.8 s, whereas human analysis showed a standard deviation of ±2,384 s, indicating marked heterogeneity among observers and across slides. This contrast is visually summarized in Figure 4, which compares analysis time and accuracy per slide between human researchers and the artificial intelligence system. The combination of higher mean accuracy, lower between-slide variation, and lower temporal dispersion suggests that the automated system is more predictable and reproducible-features that are especially desirable for large-scale screening workflows in settings with high diagnostic demand and limited workforce.

Figure 4
Comparison of time and accuracy between researchers and AI per slide.

Efficiency and effectiveness of AI-assisted diagnosis in complex scenarios

The results show that AI analysis time remained virtually unchanged regardless of microfilarial density or slide morphological complexity, ranging from 87 to 117 s. In contrast, human performance was sensitive to increasing parasite burden, with longer analysis times and greater variability. In practice, traditional microscopic diagnosis required about 30-40 minutes per slide, whereas AI-assisted diagnosis remained within the 1-2-minute range, with only minimal fluctuation.

From a health service management perspective, this means that during vector-control campaigns or periods of high demand, AI can process a substantially larger number of slides while maintaining a stable turnaround time, thereby supporting more timely public health decision-making.

Model performance in the automated classification of microfilariae

Model performance was evaluated at two complementary levels: object-level classification (individual microfilariae), used as a supportive descriptive analysis, and image-region classification (the patch-based approach), adopted as the primary evaluation aligned with the architecture’s classificatory nature used and heterogeneity observed in microscopic fields.

At the object level, of the 96 microfilariae manually identified across nine slides, the model correctly classified 92, corresponding to an identification rate of 95.8%. Missed detections were concentrated in slides with higher parasite density, where structural overlap increases optical complexity. Even under these conditions, however, performance remained high, with only limited losses and overall consistent interpretation. This finding is particularly relevant for parasitological surveillance, since false negatives pose a direct epidemiological risk by leaving infected animals outside the scope of control actions.

The primary model evaluation was conducted at the image-region level, with each patch treated as an independent unit of analysis. After segmentation into patches, 900 regions were evaluated, yielding a confusion matrix characterized by a high proportion of true negatives and a low number of false negatives, indicating stable overall performance aligned with the study aims. The predominance of true negatives suggests that the model can reliably recognize parasite-free areas, while the low false-negative rate supports its suitability for automated screening applications.

The model showed high sensitivity (95.8%), revealing a strong ability to correctly identify regions containing microfilariae even under adverse conditions such as variation in focus, illumination, and parasite burden. Moderate precision (63.4%) reflected the occurrence of false positives, mainly attributable to the morphological resemblance between microfilariae and non-parasitic artifacts. This profile is acceptable and operationally desirable in screening systems, where minimizing false negatives is prioritized, and positive classifications can subsequently be reviewed by human specialists.

The F1-score (0.763) indicated a satisfactory balance between sensitivity and precision, confirming consistent performance even in heterogeneous microscopic fields. Overall accuracy (93.7%), calculated across all analyzed regions, further demonstrates classifier stability throughout the full dataset, including both parasitized and non-parasitized areas. A quantitative summary of the performance metrics is shown in Chart 1, which consolidates the main indicators used to evaluate the model.

Chart 1
Summary of the performance results of the artificial intelligence model in patch-based classification for the detection of microfilariae in optical microscopy images.

Metric Evaluated	Value achieved	Technical Meaning	Clinical/Operational Interpretation
Sensitivity (Recall)	95.8%	Ability to correctly detect the microfilariae present	Highly sensitive: identifies almost all microfilariae, reducing the risk of false negatives and epidemiological impact.
Precision	63.4%	Ability to avoid incorrect markings	Moderate accuracy: some detections are artifacts, requiring additional visual review, but without affecting the screening workflow.
F1-Score	76.3%	Balance between recall and accuracy	Balanced performance: the F1-Score indicates a good balance between sensitivity and accuracy, suitable for automation.
Overall Accuracy	93.6%	Percentage of correct matches considering all patches analyzed	Stable classification: maintains high performance even with morphological variation and high parasite load.
AUC (ROC Curve)	94.6%	Overall capacity to discriminate between regions with and without microfilariae	High discrimination (AUC): correctly differentiates positive patches in approximately 94.6% of comparisons.
TP	92	Accurate detection of true microfilariae	Good localization capability: detects most microfilariae present in the analyzed field.
FN	4	Microfilariae not detected	Rare underdiagnosis: low false negative rate, critical for epidemiological surveillance and control.
FP	53	Non-parasitic structures classified as microfilariae	Acceptable false positives: incorrect markings only generate a recheck, preserving diagnostic accuracy.
TN	751	Areas correctly ignored	Robust negative control: accurately recognizes parasite-free areas, avoiding unnecessary alarms in large-scale analyses.

• Notes: TP = True positives; FN = False negatives; FP = False positives; TN = True negatives.
• Source: Authors (2025).

The classifier’s discriminatory ability was further assessed through the ROC curve, which showed consistent separation between positive and negative regions across different decision thresholds. The area under the curve (AUC=0.946) indicates excellent discriminatory power, suggesting that the model assigned higher probabilities to regions containing microfilariae in most comparisons. Threshold analysis showed that intermediate values maximize sensitivity while maintaining adequate control of the false-positive rate, a desirable feature for epidemiological surveillance and laboratory screening applications.

Overall, the results indicate that the model performs strongly in the automated classification of image regions according to the presence of microfilariae, particularly because of its high sensitivity and robustness to morphological and optical variation. These findings support the potential of this approach as a tool to assist parasitological screening, especially in low-infrastructure settings, not as a replacement for expert assessment but as a means of improving standardization, speed, and diagnostic capacity.

Discussion

Interobserver variability and limitations of manual diagnostic microscopy

Interobserver variability is one of the main challenges in diagnostic microscopy, especially in decentralized settings such as the Amazon, where professionals with different training and experience levels interpret parasitological slides under frequently suboptimal conditions. Even small differences in the morphological reading of microfilariae may compromise diagnostic accuracy, directly affecting clinical management and the quality of epidemiological surveillance actions, as widely described in studies of manual parasitological diagnosis²².

This dependence on individual experience may be further exacerbated by operational factors such as visual fatigue, inadequate lighting, and optical limitations of the available equipment, making manual diagnosis particularly vulnerable to error in high-demand settings with restricted infrastructure.

Diagnostic standardization and reproducibility through convolutional neural networks

Artificial intelligence based on convolutional neural networks is particularly valuable for its ability to standardize diagnostic decisions. The model proposed in this study, trained on expert-validated images and confirmed by molecular testing, showed homogeneous and reproducible behavior regardless of microscopic field complexity or image quality. This finding is consistent with evidence that CNNs can extract robust morphological patterns even from noisy and heterogeneous images^23-25.

AI-promoted standardization has direct implications for laboratory routine. By reducing the influence of subjective factors such as visual fatigue, individual experience, and operational variability, automated models contribute to more uniform results and greater diagnostic reliability. This characteristic is especially relevant in locations where specialized technical supervision is limited or unavailable, a common condition in rural and riverine areas of the Amazon²⁶.

Automated screening, human-machine cooperation, and operational efficiency

Another key finding of this study was the contribution of AI to automated screening. The system identified suspicious regions in digitized slides, directing the parasitologist’s attention to fields with greater diagnostic relevance. This approach represents a model of human-machine cooperation in which technology functions as a clinical decision-support tool, improving the efficiency and consistency of human work, aligned with contemporary principles of diagnostic support systems²⁷.

Screening automation directly improves laboratory workflow. In high-demand situations such as surveillance campaigns, outbreaks, or periods of seasonal vector increase, artificial intelligence can rapidly filter negative or low-complexity slides, allowing specialists to focus on more challenging cases. This operational gain translates into greater efficiency and more predictable turnaround times in automated diagnostic systems²⁸.

The patch-based approach proved particularly suitable for filiform and visually subtle organisms such as microfilariae because it enables local analysis of heterogeneous regions without requiring exhaustive bounding-box annotations, which are often impractical in decentralized settings with limited technological complexity.

Health equity, algorithmic justice, and applicability in vulnerable territories

From a health equity perspective, the possibility of running the model offline on low-cost equipment represents a substantial advance. The use of webcams coupled to bright-field microscopes and local processing makes the technology feasible in territories with limited connectivity, promoting the concept of algorithmic justice by expanding access to diagnosis for historically underserved populations^29,30.

This approach is aligned with the SUS Digital Health Strategy guidelines, which prioritizes scalable, sustainable technological solutions that are sensitive to territorial realities. Ordinance GM/MS No. 3,691/2024 reinforces the need to strengthen the problem-solving capacity of primary care and health surveillance through digital tools adapted to local contexts, further supporting the relevance of the proposed model. AI solutions that can be run locally on accessible hardware not only broaden access to diagnosis but may also help reduce structural health inequalities^21,31.

Computational efficiency, energy sustainability, and study limitations

Choosing efficient architectures such as EfficientNetV2-B0 was technically appropriate for settings with computational constraints. Previous studies have shown that compact convolutional networks can achieve high levels of accuracy and AUC even with low-quality images while maintaining strong operational and energy performance^32,33.

Recent studies further indicate that AI models can be deployed on mobile devices or notebooks with reduced energy consumption, expanding their applicability in field settings. Analyses of the energy efficiency of neural networks during inference show that, with optimized hardware, even deep models can operate sustainably outside traditional laboratory environments³⁴.

Although the findings are promising, the consolidation of this technological solution still depends on the expansion of the image bank and on multicenter validation. Including samples from different territories, laboratory settings, morphological variants, and other filarial species will be essential to improve the model’s generalizability. Even so, the present results indicate that lightweight, portable, territory-oriented solutions represent relevant tools for reducing long-standing inequalities in access to the diagnosis of neglected tropical diseases and for supporting the sustainable digital transformation of the SUS³⁵.

Conclusion

This study shows that applying a convolutional neural network-based artificial intelligence to identify microfilariae in blood smears is technically feasible and operationally consistent in the Amazonian context. Integrating strict biological confirmation with an optimized computational pipeline yielded high performance even under adverse field conditions, with gains in efficiency and standardization vis-à-vis conventional microscopic reading. Although additional validation is needed, the findings indicate that lightweight solutions that run on accessible hardware can serve as strategic tools to support parasitological screening and strengthen epidemiological surveillance sustainably in areas with limited infrastructure. The proposed architecture and pipeline are conceptually transferable to other filarial infections and parasitological settings, provided that they are accompanied by specific biological validation.

Acknowledgments

The authors are grateful to the Laser Microdissection Platform of Fiocruz Amazônia - Instituto Leônidas e Maria Deane (ILMD), RPT07H, for the technical support and for providing the infrastructure used in the sample preparation and microfilariae isolation stages. They also are grateful to the Library and Secretariat of ILMD for the administrative and institutional support for the development of the research, which were fundamental for the execution of the study and for obtaining the data presented here.

References

• ¹ Silva-Nunes M, Dal'Asta AP, Codeço CT. Challenges and perspectives in analyzing health in the Brazilian Amazon: a look at population-based studies. Cad Saude Publica 2025; 41(13):e00045824.
• ² Medeiros JDS, Schweickardt JC, Martins FM. Entre cheias e vazantes: uso das embarcações na produção do cuidado e acesso à saúde no território líquido em um município amazônico, Brasil. Saude Soc 2025; 33:e240381pt.
• ³ Lima RTS, Fernandes TG, Martins Júnior PJA, Portela CS, Santos Junior JDO, Schweickardt JC. Health in sight: an analysis of primary health care in riverside and rural Amazon areas. Cien Saude Colet 2021; 26(6):2053-2064.
• ⁴ Oliveira JCS, Correa NCC, Carminé RLA, Souza SES, Chui FMS. Geographic and technological challenges in the connectivity process of health care units in the rural area of Manaus-Amazon. IOSR J Bus Manag 2024; 26(1):13-18.
• ⁵ Portela CS, Suwa UF, Oliveira JCS, Simão CLG, Crainey JL. Monitoramento e controle das doenças filariais na região Amazônica Brasileira. In: Tobias R, Leles FAG, Lima MCRF, editores. Planejamento e políticas de saúde na Amazônia: fundamentos e caminhos. Rio de Janeiro: Editora Rede UNIDA; 2024. p. 121-135.
• ⁶ Savedra ASS, Rodrigues LOVSB, Oliveira WEA, Moraes DDP, Silva AM. Eficácia de intervenções para o controle de doenças tropicais negligenciadas na população infantil da Amazônia Brasileira: uma revisão sistemática. Braz J Implantol Health Sci 2025; 7(11):259-269.
• ⁷ Gordon RM, Young CJ. Parasites in dogs and cats in Amazonas. Ann Trop Med Parasitol 1922; 16(3):297-300.
• ⁸ Barbosa UC, Nava AFD, Ferreira Neto JV, Dias CA, Silva VC, Mesquita HG, Sampaio RTM, Barros WG, Farias ES, Silva TRRD, Crainey JL, Tadei WP, Koolen HHF, Pessoa FAC. Dirofilaria immitis is endemic in rural areas of the Brazilian Amazonas state capital, Manaus. Rev Bras Parasitol Vet 2023; 32(2):e000223.
• ⁹ Genchi C, Kramer L. Subcutaneous dirofilariosis (Dirofilaria repens): an infection spreading throughout the Old World. Parasites Vectors 2017; 10(Supl. 2):517.
• ¹⁰ Simón F, Siles-Lucas M, Morchón R, González-Miguel J, Mellado E, Carretón E, Montoya-Alonso JU. Human and animal dirofilariasis: the emergence of a zoonotic mosaic. Clin Microbiol Rev 2012; 25(3):507-544.
• ¹¹ Genchi C, Kramer LH. The prevalence of Dirofilaria immitis and D. repens in the Old World. Vet Parasitol 2020; 280:108995.
• ¹² Rena O, Leutner M, Casadio C. Human pulmonary dirofilariasis: uncommon cause of pulmonary coin-lesion. Eur J Cardiothorac Surg 2002; 22(1):157-159.
• ¹³ Barbosa LLS. Atlas de citologia de líquor: um guia prático para apoio médico-laboratorial em neurodiagnósticos [eBook]. São Paulo: Editora Dialética; 2025 [acessado 2026 jan 3]. Disponível em: https://www.bdtd.uerj.br:8443/handle/1/24835.
• ¹⁴ Zhang C, Jiang H, Jiang H, Xi H, Chen B, Liu Y, Juhas M, Li J, Zhang Y. Deep learning for microscopic examination of protozoan parasites. Comput Struct Biotechnol J 2022; 20:1036-1043.
• ¹⁵ Zedda L, Loddo A, Di Ruberto C. YOLO-PAM: parasite-attention-based model for efficient malaria detection. J Imaging 2023; 9(12):266.
• ¹⁶ Kumar A, Sharma S, Rizvi STH, Ali J. UltraLightSqueezeNet: ultra lightweight CNN model for malaria detection using blood smear images. arXiv 2025. Doi: https://doi.org/10.48550/arXiv.2501.14172
  » https://doi.org/10.48550/arXiv.2501.14172
• ¹⁷ Hoyos K, Hoyos W. Supporting Malaria Diagnosis Using Deep Learning and Data Augmentation. Diagnostics 2024; 14(7):690.
• ¹⁸ Mathur A, Singh P, Gupta R. Deep learning method for malaria parasite evaluation from blood smears. Artif Intell Med 2024; 145:102739.
• ¹⁹ Kalyan DS, Manogaran G, Chang CY. Automated detection of malaria parasites using deep learning models: a review. Inform Med Unlocked 2024; 57:101570.
• ²⁰ Fan BE, Yong BSJ, Li R, Wang SSY, Aw MYN, Chia MF, Chen DTY, Neo YS, Occhipinti B, Ling RR, Ramanathan K, Ong YX, Lim KGE, Wong WYK, Lim SP, Latiff STBA, Shanmugam H, Wong MS, Ponnudurai K, Winkler S. From microscope to micropixels: a rapid review of artificial intelligence for the peripheral blood film. Blood Rev 2024; 64:101144.
• ²¹ Brasil. Ministério da Saúde (MS). Portaria GM/MS nº 3.691, de 23 de maio de 2024. Institui diretrizes para a transformação digital da saúde no âmbito do SUS. Diário Oficial da União; 2024.
• ²² Rajaramam S, Antani SA, Catala A. Artificial intelligence in digital pathology: automated detection of blood cell parasites. Front Microbiol 2019; 10:2732.
• ²³ Rajaramam S, Antani SK, Poostchi M, Silamut K, Hossain MA, Maude RJ, Jaeger S, Thoma GR. Pre-trained convolutional neural networks as feature extractors toward improved malaria parasite detection in thin blood smear images. PeerJ 2018; 6:e4568.
• ²⁴ Rajaramam S, Jaeger S, Antani SK. Performance evaluation of deep neural ensembles toward malaria parasite detection in thin-blood smear images. PeerJ 2019; 7:e6977.
• ²⁵ Fuhad KMF, Tuba JF, Sarker MRA, Momen S, Mohammed N, Rahman T. Deep learning based automatic malaria parasite detection from blood smear and its smartphone based application. Diagnostics (Basel) 2020; 10(5):329.
• ²⁶ Kalyan DS, Manogaran G, Chang CY. Automated detection of malaria parasites using deep learning models: a review. Inform Med Unlocked 2024; 57:101570.
• ²⁷ Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 2019; 25(1):44-56.
• ²⁸ Olawade DB, Fidelis SC, Marinze S, Egbon E, Osunmakinde A, Osborne A. Artificial intelligence in clinical trials: a comprehensive review of opportunities, challenges, and future directions. Int J Med Inform 2026; 206:106141.
• ²⁹ Bahr J, Suutala A, Diwan V, Mårtensson A, Lundin J, Linder N. AI-supported digital microscopy diagnostics in primary health care laboratories: scoping review. J Med Internet Res 2026; 28:e78500.
• ³⁰ Pinto-Coelho L. Como a inteligência artificial está moldando a tecnologia de imagem médica: uma análise de inovações e aplicações. Bioengineering 2023; 10(12):1435.
• ³¹ Brasil. Ministério da Saúde (MS). Estratégia de Saúde Digital para o Brasil 2020-2028. Brasília: MS; 2020.
• ³² Mujahid M, Rustam F, Shafique R, Montero EC, Alvarado ES, de la Torre Diez I, Ashraf I. Efficient deep learning-based approach for malaria detection using red blood cell smears. Sci Rep 2024; 14:13249.
• ³³ Montalbo FJP, Alon AS. Empirical analysis of a fine-tuned deep convolutional model in classifying and detecting malaria parasites from blood smears. KSII Trans Internet Inf Syst 2021; 15(1):147-165.
• ³⁴ Desislavov R, Martínez-Plumed F, Hernández-Orallo J. Trends in AI inference energy consumption: beyond the performance-vs-parameter laws of deep learning. Sustain Comput Inform Syst 2023; 38:100857.
• ³⁵ Bahr J, Diwan V, Mårtensson A, Linder N, Lundin J. AI-supported digital microscopy diagnostics in primary health care laboratories: protocol for a scoping review. JMIR Res Protoc 2024; 13:e58149.

• Funding
  This study was funded by the Conselho Nacional de Desenvolvimento Cientifico e Tecnologico (CNPq) and the Fundação de Amparo à Pesquisa do Estado do Amazonas (FAPEAM), which supported the development of the research by granting scholarships to researchers and promoting scientific research activities, contributing to the execution of the methodological steps, the analysis of the data and the consolidation of the results presented.
• Research ethics committee
  This project was submitted to and approved by two Research Ethics Committees, in accordance with the ethical principles established in Resolution No. 466/2012 of the National Health Council (CNS). The first opinion was issued by the Research Ethics Committee of the Dr. Heitor Vieira Dourado Tropical Medicine Foundation (CEP/FMT-HVD), under the Certificate of Presentation of Ethical Appraisal (CAAE) No. 73732523.7.0000.0005. Subsequently, the research was approved by the Research Ethics Committee of the Northern University Center (CEP/UNINORTE), under CAAE No. 70265923.9.0000.0010 and opinion No. 6.179.007.
• Data availability statement
  The data sources adopted in the research are indicated in the article’s body.

• Chief editors:
  Maria Cecília de Souza Minayo, Romeu Gomes, Antônio Augusto Moura da Silva, Vania de Matos Fonseca

Data availability

The data sources adopted in the research are indicated in the article’s body.

Publication Dates

• Publication in this collection
  29 June 2026
• Date of issue
  May 2026

History

• Received
  05 May 2025
• Accepted
  26 Jan 2026
• Published
  28 Jan 2026