Ethical statement

Ethical approval was obtained from the Comité Bioética de la Facultad de Medicina (approved on 30/10/2020), Universidad Mayor de San Simón, Bolivia and from the Research Ethics Committee (REC) Instituto de Salud de Carlos III, Spain (CEI PI 74_2020).

Overall workflow

This study details a comprehensive workflow for developing a computer-assisted diagnostic tool, from sample digitization to on-device deployment (Fig 1). The workflow begins with sample collection and digitization using a smartphone based system. The digitization process, detailed previously [29,30], involves attaching the smartphone to the ocular of a conventional optical microscope via a 3D-printed adapter. A custom Android application enables users to digitize the microscopy field of view and upload the images to a telemedicine platform for expert annotation, placing a bounding box around each identified parasite, thereby creating the necessary training dataset. To enable parasite localization and quantification, we selected AI models for object detection as the core methodology. Accordingly, we trained and evaluated three distinct object detection models: SSD-MobileNetV2, YOLOv8-Small, and Faster R-CNN with a ResNet-50 backbone. Finally, the best-performing model was optimized for deployment on smartphones without requiring internet connectivity to assist clinicians in the diagnostic process in real time.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Simplified study design and workflow.

https://doi.org/10.1371/journal.pntd.0012955.g001

Dataset

Our dataset comprises two subsets: we created a new dataset (Laboratorios de Investigación Médica (LABIMED) dataset) consisting of human thick and thin blood smears and cerebrospinal fluid smears (CSF), and a publicly available dataset containing mice thin blood smears (Fig 2). Integrating these diverse samples exposed the model to greater clinical and experimental variability. This process was crucial for improving the model’s adaptability, as sample preparation, staining protocols, and imaging conditions often differ between laboratories. For instance, human thick blood smears retain multiple blood cell layers, whereas thin smears provide single-cell layers, altering parasite morphology and background noise. Similarly, staining methods affect color contrast and parasite visibility. Variations in imaging equipment - such as microscope resolution, lighting conditions, or camera - further compound these discrepancies.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Data preprocessing.

(A) Field of view image, digitized with smartphone. (B) Mosaic augmentation example.

https://doi.org/10.1371/journal.pntd.0012955.g002

A total of 36 smear preparations were collected from LABIMED, a reference laboratory for parasitology in Cochabamba, Bolivia, involving 9 positive subjects. Among these, 8 provided blood smear samples and 1 provided a cerebrospinal fluid (CSF) sample. All samples were stained with Giemsa. Specifically, 9 preparations from a single subject were CSF samples, 13 preparations from 4 subjects were thin blood smears, and 14 preparations from another 4 subjects were thick blood smears. The samples were digitized using a Samsung Galaxy J7 prime smartphone coupled to an Olympus CH30 microscope. Images were acquired through a 100x oil immersion objective with a numerical aperture of 1.25, creating final images of 4128x3096 pixels. Digitization was made using a 3D printed adapter coupled into the microscope, allowing the use of conventional microscopes to collect the images. Both positive and negative images were captured, resulting in 726 images.

The public dataset comprises 33 mice thin blood smear samples, totaling 675 images [31]. All images were Giemsa-stained and digitized using a smartphone camera coupled to an optical microscope with a 100x objective. Most images were captured at a resolution of 3,456 × 4,608 pixels, though a subset (2,448 × 3,264 pixels) reflects adjustments in camera settings during acquisition.

Image labeling

For the development and evaluation of our object detection algorithm for Trypanosoma cruzi, accurate image labeling was crucial. We started by labeling the LABIMED dataset from scratch. Bounding boxes were placed around the identified parasite, indicating the location and the class of the parasite. In addition, some artifacts that look similar to the parasite were annotated. A total of 931 parasites were identified and annotated. Then, we trained a preliminary model to generate predictions on Morais et al. dataset, these predictions were subsequently reviewed and refined by a human annotator, who adjusted the bounding boxes to ensure they accurately aligned with the parasites and corrected any false positives. This iterative process led to the creation of 3,151 labeled objects.

The labeled dataset was carefully split into two independent sets at sample preparation level for LABIMED dataset. This approach ensured that all images from the same preparation belonged to the same dataset, whether used for training the AI model or validating its performance. The split was carried out after labeling the images to guarantee that each sample preparation type is presented in both the training and validation sets. The split was performed randomly, striving to achieve an 80%-20% split between the two sets. The result distribution is presented in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Dataset distribution for model training and validation.

https://doi.org/10.1371/journal.pntd.0012955.t001

Data preprocessing

Given the alignment of the smartphone with the microscope eyepiece, the area captured by the mobile phone was limited to a circular region, as illustrated in Fig 3. As the parasite size is very small, typically occupying an area of only 100x100 pixels within an image that is 4128x3096 pixels in size, to exclude non-informative regions, such as the black areas outside the circular field of view, we opted to use square images. The size of the field of view is approximately 2800x2800 pixels, and the inner square is approximately 1900 × 1900 pixels. To increase the relative size of the parasites within the image, we selected square images that encompass the entire field of view, as represented by the white square in Fig 2. This approach ensured that the visualized area is maximized while focusing on the relevant regions for analysis.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Microscopic images of blood and cerebrospinal fluid samples highlighting field of view and parasite visualization.

The left panels display the full field of view for each sample type, while the right panels provide close-up views of parasites identified within the samples. Specifically, A’ shows a parasite from the thin blood smear (Panel A), B’ depicts a parasite from the thick blood smear (Panel B), C’ presents a parasite from the mice thin blood smear (Panel C), and D’ highlights a parasite from the cerebrospinal fluid smear (Panel D).

https://doi.org/10.1371/journal.pntd.0012955.g003

In addition, to overcome the problem of imbalance foreground and background in object detection tasks, we manually increased the number of parasites by applying mosaic augmentation, which consisted of cropping a 160x160 pixel patches that contained at least one parasite, and blending 16 images to create a new image of 640 x 640 pixel. This augmentation was only applied for the training set.

AI algorithm

For the automated analysis of microscopy images, our primary goal was not only to confirm the presence of Trypanosoma cruzi but also to locate and count each individual parasite. Simple image classification models, which only provide a label for an entire image (e.g., “parasite present”), are insufficient for this task as they do not provide location or quantity. Therefore, we chose an object detection approach. Object detection models are specifically designed to perform both classification (identifying an object as T. cruzi) and localization (placing a bounding box around each instance). This dual capability allows for the precise identification and enumeration of parasites within a given microscopic field, making it the ideal methodology for our study.

Object detection models can be divided into two groups: single stage detectors and two stage detectors. On the one hand, single stage detection algorithms are designed to make predictions within a single pass over the image which makes them faster but less accurate since they do not use a region proposal stage. On the other hand, two stage detection algorithms are more complex since they have a module that uses a CNN to propose regions of interest that are later refined in a second stage that performs the classification and localization. This process is usually more accurate but due to their nature, they are slower and require more computational resources.

In our work we focused on optimizing single-stage detectors to balance the need for speed and accuracy- as our goal was for the AI model to be run in a mid-range Android device in real time. This approach ensures that real-time detection capabilities are maintained. In particular, we evaluated two different algorithms that represent the two main approaches for single stage detectors: Single Shot Detection (SSD) and YOLO [39–41].

The SSD-MobileNetV2 model integrated the SSD framework with the lightweight MobileNetV2 architecture. SSD predicted bounding boxes and class probabilities directly from feature maps at different scales in a single pass, enabling efficient object detection. MobileNetV2, designed for mobile optimization, employs depthwise separable convolutions to minimize computational costs, making it ideal for deployment on resource-constrained devices like smartphones. This combination struck a balance between speed and accuracy. We initialized the model using pretrained weights from the COCO image dataset. For optimizing the loss function, we employed the RMSprop optimizer with an exponential decay learning rate, starting with a base learning rate of 0.008.

Conversely, YOLOv8 is one of the latest advancements in the YOLO series. It utilizes CSPDarknet53 as its backbone for feature extraction and incorporates a path aggregation network to enhance information flow across scales. YOLOv8 also introduces an anchor-free detection mechanism that predicts object centers directly, reducing the need for numerous box predictions and thus accelerating detection. The YOLOv8 family includes various variants, such as Nano, Small, Large, and Extra Large, each balancing accuracy and inference speed. For our application, we selected YOLOv8-Small for its optimal balance of these factors. The models were trained using the AdamW optimizer, configured with a learning rate of 0.002, a first momentum parameter of 0.9, and a second momentum parameter of 0.999. We also initialized the model using pretrained weights from the COCO image dataset for comparison purposes.

For both models, input size of 640x640 was used. During training, both models were subjected to on-the-fly data augmentation. These augmentations included random flips, rotations with a 50% probability, and random crops that maintained at least 80% of the original image area. We also implemented color jittering, randomly adjusting the image’s hue by a factor of ±0.015, saturation by a factor of 0.7, and value by a factor of 0.4. Additionally, early stopping with 10 epochs patience was employed to prevent overfitting, ensuring robust model performance.

For comparison purposes, we also trained a two-stage detector, Faster RCNN even though it is not suitable for smartphone deployment due to its higher computational demands. Faster R-CNN is a two-stage object detection model with high localization and classification accuracy. Our implementation utilized a ResNet-50 backbone to extract deep feature maps from the input images. Its first stage, the Region Proposal Network, identifies candidate object regions. These proposals are then passed to the second stage, where the model performs final classification and precise bounding box regression. This two-stage approach, while more computationally intensive than single-stage methods, allows the model to achieve a high degree of precision. For fair comparison, we also initialized this model using pretrained weights from the COCO dataset and used the RMSprop optimizer with a learning rate of 0.01 and momentum of 0.9.

We used two primary frameworks for model training to leverage the optimal tools for each architecture. The SSD-MobileNetV2 and Faster R-CNN models were trained using the TensorFlow Object Detection API. The YOLOv8-Small model was trained using the official Ultralytics framework, which provides state-of-the-art implementations specifically for the YOLO family of models. All training was conducted on the Amazon SageMaker platform, utilizing an NVIDIA A10G GPU with 16 GB of memory and a Thermal Design Power (TDP) of 150 W. The total training time for the project was approximately 42 hours, resulting in an estimated carbon footprint of 2.43 kg CO₂ eq.

AI model preparation for deployment

The trained AI model was exported to tflite format, with a size of approximately 11 megabytes for SSD Mobilenet and 45MB for YOLOv8 model. These models can be executed on Smartphones in real time. In addition, we created a customized Android application to execute the model. This app has the key requirement of working in real time without internet connectivity in order to be suitable for work in regions with intermittent or unavailable connectivity and/or electricity.

Model performance on cross-validation and random split evaluations

Due to the limited sample size, we employed a two evaluation strategy to assess model performance. First, we performed five-fold cross-validation for each model to obtain a more reliable estimate while mitigating fold-to-fold variability. In parallel, we trained a version of each model using a random 80/20% split of the data to evaluate performance in a conventional training-testing scenario.

Following data preprocessing, the final training set comprised 7,249 augmented images with 66,557 annotated parasites, while the validation set, used to assess model generalizability, included 1,099 images containing 1,668 labeled targets. After training, the top-performing model checkpoint (selected based on validation performance) was used to analyze validation images, successfully identifying T. cruzi parasites in diverse microscopy samples as visualized in Fig 3.

To evaluate the performance of our models, we used precision, recall, and F1 score as metrics. For our evaluation, a detection was classified as a true positive if its confidence score exceeded the threshold and its Intersection over Union (IoU) with a ground truth box was greater than 0.3. A detection was classified as a false positive if it did not overlap with any ground truth box. Finally, a ground truth object was classified as a false negative if it was not detected by the algorithm. The threshold value was chosen to balance precision and recall, aiming to detect as many parasites as possible while limiting the number of false positives. If the priority is to ensure no parasites are missed, a lower threshold can be used to increase sensitivity. Conversely, to minimize false positives, a higher threshold can be applied. Precision, recall, and F1 score are then defined as follows:

Cross-validation across five folds revealed subtle differences in the performance of the three detection models. FasterRCNN-ResNet50 maintained consistently high F1-scores (85.8–94.5%), with slightly higher recall in fold 4 (93.6%) and lower in fold 2 (89.3%). SSD-Mobilenet v2 showed greater variability, particularly in recall (85.8–94.5%), achieving its highest F1-score in fold 4 (90.7%). YOLOv8-small generally had higher precision (91.3–92.8%) but lower recall (86.0–89.3%), resulting in moderate F1-scores (88.8–91.0%). Overall, while FasterRCNN-ResNet50 provided the most balanced and consistently high performance, the differences among the models were relatively small, suggesting all three architectures are competitive, with FasterRCNN offering a slight edge in reliability across folds (Table 2).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Cross-validation performance summary (mean ± SD, %).

https://doi.org/10.1371/journal.pntd.0012955.t002

Building on these results, we further evaluated model performance using a random 80/20% data split to provide an additional, complementary assessment of model robustness (Table 3) SSD-MobileNetV2 demonstrated superior performance on human samples, achieving a precision of 86%, a recall of 87%, and an F1 score of 86.5%. In comparison, YOLOv8-Small recorded precision, recall, and F1 scores of 83.6%, 87%, and 80.5%, respectively. Conversely, YOLOv8-Small outperformed SSD-MobileNetV2 on mice samples, attaining precision, recall, and F1 scores of 93.7%, 91.3%, and 92.5%, while SSD-MobileNetV2 achieved 95.8%, 85%, and 90.1%, respectively. Notably, the two-stage detector Faster R-CNN exhibited higher performance than the single-stage detectors, with precision, recall, and F1 scores of 93.6%, 81.4%, and 87.1% on human samples, and 96.8%, 90.3%, and 93.5% on mice samples. These results complement the cross-validation findings, highlighting Faster R-CNN’s robustness and balanced performance across sample types.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Model performance on human and mice samples, and the inference time measured on an Oppo Reno 6 smartphone.

https://doi.org/10.1371/journal.pntd.0012955.t003

The reported inference time measures the average duration to process a single image frame on our target hardware. This measurement begins when an image is received and ends when the final predictions are returned. It therefore encompasses all necessary steps: image pre-processing (e.g., resizing and normalization), the model prediction itself using the TensorFlow Lite interpreter, and post-processing to decode the output into usable bounding boxes.

When analyzing performance across different sample types, the models showed optimal results for mice with thin blood smears, achieving a precision of 95.8%, a recall of 85%, and an F1 score of 90.1%. In contrast, the performance on human thick blood smears was lower, with precision, recall, and F1 scores of 69.2%, 64.3%, and 66.7%, respectively. Performance for human thick smears drops compared to human CSF, thin, or mouse smears due to fewer training images and labels. These detailed results are presented in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Performance of SSD MobileNet v2 separated by sample type.

https://doi.org/10.1371/journal.pntd.0012955.t004

Pilot experiment of the AI-Augmented microscopy system in a laboratory environment

To complement the controlled cross-validation and random-split experiments, we further assessed the AI algorithm in an experiment in real-world conditions using independently collected blood smears and a smartphone-based scanning workflow (Fig 4). This approach allowed us to assess how the models perform outside of curated datasets in a lab environment capturing practical challenges including user interaction. By benchmarking the algorithm against expert annotations in a real-time scanning scenario, we were able to directly evaluate its robustness in conditions that closely mimic field deployment.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Operational deployment of real-time AI Algorithm.

(A) Setup for image acquisition and real-time processing using a smartphone coupled to a light microscope. (B) Screenshot illustrating the real-time AI detection of parasites on the smartphone interface.

https://doi.org/10.1371/journal.pntd.0012955.g004

Three independent thin blood smears were used to assess the AI algorithm in real time in a lab environment. The evaluation protocol required a user to scan the smears using a smartphone running the algorithm in real-time. Images were saved under two conditions: (1) when the AI detected a potential parasite (true or false positive) and (2) when the user manually identified a parasite missed by the AI (false negative).

This approach of real-time AI augmented microscopy recorded the key fields of view, A single 15 × 15 cm scan at 100 × magnification can generate over 500,000 fields of view, making recording of every frame impractical unless a mechanical scanner is used. Consequently, only images containing potential findings were saved, while correctly identified negative fields (true negatives) were not recorded. All captured images and corresponding detections were then uploaded to the telemedicine platform for independent review by two experts: Expert A (senior) and Expert B (junior).

Performance results

Inter-rater reliability between the two experts was strong, with a Cohen’s Kappa score of 0.868. This high agreement established confidence in the annotated ground truth, against which the AI algorithm was benchmarked (Expert A designated as the ground truth).

At the object level, the algorithm achieved a precision of 67.1%, a recall of 96.4, resulting in an F1-score of 79.1% (Fig 2). This performance profile suggests the system configuration is well-suited for screening applications.

This real-time evaluation showed the system’s operational practicality and the value of human oversight. The algorithm generated an average of seven false positive detections per scanned sample, a manageable rate of false alarms that can be efficiently reviewed and dismissed by the human user. The human operator also recovered the 3.6% of parasites missed by the algorithm detections among the recorded fields of view. This finding highlights how the human operator effectively leveraged the AI’s help with detections, while compensating for the AI’s misses, minimizing the likelihood of a critical error and affirming the value of the human-in-the-loop approach.