2.2.1 Electronic data capture.

To enable a fully digital, end-to-end workflow, the AI-DP platform assigned a unique QR code label to every sample container, slide, and result form. A mobile EDC app scanned these labels (Fig 3) and automatically linked each sample identifier to timestamps (collection, preparation, and slide-reading), study site registration details and participant demographics, geospatial coordinates, and optionally water, sanitation, and hygiene (WASH) indicators. Only pseudonymised identifiers (e.g., study‑specific participant IDs) were stored digitally; no direct personal identifiers were captured within the EDC application or scanner software. Although the EDC application supports encryption of fields containing personally identifiable information at rest, this functionality was not used in the present study because no such identifiers were collected digitally.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Electronic data capture (EDC) application and reporting dashboard.

Panel A: customisable main screen of the mobile EDC app for registering participants and study sites, and collecting metadata such as human-microscopy Kato-Katz egg counts. Panel B: built-in QR-code scanner to timestamp events such as slide-preparation. Panel C: geolocation interface for mapping sample-collection sites, with adjustable pin and latitude/longitude display (map base layer: Natural Earth 1:10m Admin 0 – Countries; public domain). Panel D: web-based dashboard presenting real-time data summaries, prevalence charts and overall study-progress metrics, implemented using Apache Superset (Apache‑2.0 license) and displaying geographic distributions using Natural Earth country boundaries (public domain). Panel E: detailed verification panel showing per-scan status (verified, reviewed, unverified) alongside egg-count results in tabular form, implemented using Apache Superset (Apache‑2.0 license).

https://doi.org/10.1371/journal.pntd.0013432.g003

Data could be captured offline and later synchronised to a central or cloud‑based repository, and exchanged with established EDC systems, such as Open Data Kit (ODK) and District Health Information System 2 (DHIS2), via application programming interfaces (APIs), ensuring interoperability with national health information systems. The linkage key mapping participant IDs to personal information was maintained by local study managers on paper in a secure location in-country.

2.2.2 Data storage, access control, and transmission.

During field operation, all data generated by the AI‑DP platform, including whole‑slide images, AI‑generated detections, verification results, and associated metadata, were stored locally on the WSI scanners and/or the Slide Manager. The platform was designed to operate fully offline, and no data transfer is required for image acquisition, AI inference, or result verification.

Access to locally stored data was restricted through operating‑system and application‑level controls, including authenticated access to databases and verification tools. Personally identifiable information was not stored digitally within the platform; linkage keys mapping study‑specific identifiers to participant identities were maintained separately by local study teams in accordance with ethical approvals.

When connectivity was available, selected data and results could be transmitted between scanners, mobile devices, and cloud‑based dashboards using encrypted communication protocols. Upload of image data was optional and user‑controlled. In low‑connectivity settings, data can alternatively be exported to external storage for later synchronisation.

2.2.3 Whole-slide image scanner.

We designed the custom WSI scanner for two-button operation (‘Load slide’ and ‘Start scan’). Standard 75 x 25 mm glass slides were clamped into a magnetic holder on a motorised XY stage. A secondary overview camera read the slide’s QR code, confirmed the sample type and outline of the KK thick smear, and presented the scan region to the user for confirmation. Operators could adjust the scanning region manually if required and then proceed to start the scan. During scanning, the stage advanced the slide through the region while a high-resolution camera captured images at every field of view within the determined boundary.

Imaging was performed with a 1/2.6-inch, 2-megapixel CMOS sensor (3x3 μm pixels) coupled to an inverted lens (7.5 mm focal length, 32.5 mm back focal length) yielding 3.2x magnification, a 1,280 x 720 μm field of view, a single pixel resolution of 0.93 μm, and a theoretical optical resolution of 1.2 μm. This provided an optical resolution comparable to traditional brightfield microscopy with a 10x/0.25NA Plan Achromat objective. The camera streams at 120 frames per second with illumination being provided by an LED source and diffuser. Captured images could be digitally enlarged via the verification tool.

KK thick smears typically spanned 20 mm in diameter and 135 μm in thickness (from 41.7 mg stool) with local thickness variations. To ensure at least one in-focus image per field of view, the scanner captured Z-stacks of 8 images at 20 μm intervals (total depth 240 μm) using a liquid lens (having ±0.67 mm range and 2.8 μm step size). A Sobel-filter focus metric ranked each stack; the single sharpest image was processed by the AI model for helminth egg detection and classification. Users could also perform manual refocusing post-scan, analogous to standard microscope workflows.

The WSI scanners were powered by direct current (DC) 19 V from common laptop chargers for continuous operation, or off a 19V 27,000 mAh USB power bank for up to 3 hours. The WSI scanners included a 1 TB drive permitting approximately 1,000 scanned slides. On-board AI inference ran on an NVIDIA Jetson Xavier NX in a standalone setup or on an NVIDIA Jetson Nano in a high throughput scenario when paired with a Slide Manager (discussed further under AI development).

2.2.4 AI result generation.

Image processing ran either on-board each WSI scanner in standalone mode or centrally on a Slide Manager when multiple scanners were deployed. The Slide Manager, powered by an NVIDIA Jetson AGX Xavier, provided substantially more compute than the Jetson Nano modules in individual scanners. In centralised operation, images from up to five scanners could stream wirelessly to the Slide Manager, which performed AI inference, consolidated storage, and delivered near-real-time results. This architecture simplified data management, quality control, and remote verification. During development tests, five scanners ran in parallel under one Slide Manager, while in field deployments in Uganda and Ethiopia, we routinely paired two scanners with a single Slide Manager.

During scanning, each field of view image was sent to the Slide Manager, where an offline AI model automatically detected, classified, and quantified helminth eggs as images were acquired. We followed the KK thick smear protocol to convert raw egg counts to eggs per gram (EPG) [5]. Field of view images included a 10% overlap (≈100 µm) in both axes to ensure full smear coverage and reduce edge artifacts. To prevent double-counting, we grouped duplicate detections across overlapping fields using k-means clustering based on spatial proximity and type similarity. We then categorised EPG values as negative, light, moderate, or heavy infection according to WHO thresholds [23]. Per species results were available immediately in the local interface and, when internet connectivity is available, can be optionally uploaded, with or without accompanying images to a cloud-based dashboard for further analysis and remote verification.

2.2.5 Result verification and data annotation tool.

To support diagnostic validation and continuous AI improvement, we developed EggInspector, an application that allowed for human-in-the-loop verification and object annotation. The application was available in both offline (local field) and online (cloud-synchronised) modes. On launch (Fig 4A), users choose between local data or previously synchronised studies. The Browse tab (Fig 4B) lists scanned slides, their fields of view, and detection filters (by helminth species or verification status). In the ‘Verify Objects’ view (Fig 4C), a slide overview showed AI detection overlays alongside a species-specific list of candidate helminth eggs. Verification proceeded in a 3x3 grid of cropped thumbnails (Fig 4D), one species at a time; users may accept, reject, or reclassify each detection. Selecting any thumbnail opened the full field of view image (Fig 4E), highlights all detections, and enabled focal-plane adjustment via the captured z-stack.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. EggInspector, a human-in-the-loop AI verification and image annotation tool.

Panel A: ‘Browse’ page lists scanned slides with selectable fields of view, detection-status filters (by species or verification status) and table of candidate detections. Panel B: ‘Verify Objects’ page displaying a selected scan with options to filter objects for verification, and a summary of the verification status. Panel C: batch verification grid (3x3 cropped thumbnails) of unverified Ascaris lumbricoides eggs, with Accept/Reject/Reclassify controls. Panel D: batch verification grid of verified hookworm eggs. Panel E: Contextual view of a selected hookworm egg within the full field of view, with all other detected helminth eggs highlighted and controls for z-stack focus adjustment and image adjustments. Panel F: summary page reporting total egg counts by species, eggs per gram (EPG) and options to export results locally or via the central dashboard.

https://doi.org/10.1371/journal.pntd.0013432.g004

Once verification is complete, EggInspector allowed users to aggregate species-level egg counts and export summaries to local storage or via cloud upload (Fig 4F). This hybrid AI-technician, human-in-the-loop workflow is expected to enhance diagnostic specificity by enabling users to visually verify AI-detected positives and eliminate false positives, thereby increasing confidence when ‘ruling in’ the presence of eggs. Optional manual annotation of missed eggs during model-training sessions can further increase diagnostic sensitivity by reducing false negatives.

2.2.6 Data verification and reporting.

After technician verification via EggInspector, per-species egg counts were generated and converted to EPG using the standard KK thick smear protocol [5]. Light, moderate, or heavy infection categories per species were assigned using WHO thresholds [22]. Data exports included verified egg counts by species, slide-level EPG values, intensity thresholds, associated images, and metadata. Exports could be stored locally, or when internet access was available, automatically synchronised to a cloud-based dashboard. This setup enabled near-real-time visualisation of study progress, and facilitated secondary remote verification by expert reviewers. In low-connectivity settings, completed scan datasets were exported to external storage devices (e.g., USB drives) and uploaded to central servers once a connection was established.

2.3.1 Scanning of KK thick smears.

Stool samples were collected in two NTD endemic countries: Ethiopia (Jimma town) and Uganda (West Central and West Nile regions). In Ethiopia, we recruited students aged 5–15 years (balanced by sex) from eleven primary schools. In Uganda, we enrolled schoolchildren (5–14 years) and community members aged ≥6 months. Each participant provided one stool sample in a labelled container. Samples from Jimma were transported at ambient temperature to the university laboratory; in Uganda, preparation of the KK thick smears, microscopy, and slide scanning were performed onsite under zero-infrastructure conditions (typically no running water, electricity, or internet).

Stool specimens were processed using the WHO KK method [5]. In brief, approximately 41.7 mg of sieved stool was transferred through a standard template onto a glass slide, covered with a glycerol–malachite green–soaked cellophane strip, and pressed flat. Each slide received a preprinted QR label encoding slide ID (unique pseudonymised identifier mapped to participants in a given study, a key of the mappings was stored on paper and managed by the local study coordinators), sample type ID (KK), study ID, date, and time stamp for automated tracking by the WSI scanner.

To build the ground‑truth dataset for AI training, we purposely scanned only KK slides confirmed positive for STH or intestinal SCH eggs by conventional microscopy. This approach minimised interruption of ongoing surveys and reduced data annotation effort while still providing large egg-free areas and common artefacts within each slide. Negative-control slides for estimating slide‑level specificity in diagnostic comparisons will be included in later validation studies described in [20]. Slides were typically scanned 30–60 minutes after preparation to capture hookworm eggs before degeneration. As manual microscopy within the survey took priority, some slides were scanned up to 24 hours later for improved S. mansoni egg visibility [5].

2.3.2 Ground truth dataset creation.

We assembled a high-quality WSI dataset with image tiles organised by slide ID. Each image record included slide and image IDs, species class labels and bounding box coordinates (x_min, y_min, x_max, y_max). Annotation proceeded in three phases to ensure label accuracy: (i) eight parasitologists independently drew bounding boxes and assigned helminth species labels in EggInspector; (ii) the same parasitologists then provided a second review for images they had not previously seen during their first round of annotation. During this review, verifiers could do one of the following:

• accept the image, confirming all visible labels and acknowledging that no visible eggs were missing,
• change visible objects to an ‘Unsure’ label if uncertain, reject individual objects if confident they were not helminth eggs,
• reassign a label to another helminth type (e.g., Hymenolepis, Taenia, Fasciola, or an ‘Other’ category for unspecified helminths), while identifying and annotating any omissions.

In case of disagreements, (iii) a third parasitologist adjudicated any disagreements to reach a consensus by reviewing only the objects in conflict. Here too, the adjudicator could accept, reject, mark the object as unsure, or change the label type. Images with unresolved conflicts, meaning no majority consensus was achieved on the object label even after review by multiple verifiers (e.g., differing opinions on species classification or whether an object is an egg), were excluded from the ground truth dataset until a conflict resolution process, such as a majority vote on the label, could be completed. Only images verified by at least two parasitologists with agreement on the labelled objects were retained for training, ensuring reliability and consistency in the dataset.

2.3.3 Model training and selection.

To prevent data leakage, we employed a grouped five-fold cross-validation strategy (implemented with scikit-learn’s GroupKFold [23]), ensuring all images from the same slide (and participant) were assigned to a single fold. Each fold was trained on four partitions and evaluated on the held-out partition. We applied transfer learning with the pretrained YOLOv8n model (COCO weights) [24], selecting the nano variant for its compact size and fast inference capabilities on embedded hardware. Training was initialised for 100 epochs per fold on a Tesla T4 graphics processing unit (GPU, 15 GiB) using Python 3.10.13 and PyTorch 2.1.2, using the default Ultralytics YOLOv8 hyperparameters with early stopping to prevent overfitting. The epoch with the highest Ultralytics evaluation score was selected as the best model for each fold. The code and referenced image dataset used in this study are available via Kaggle [25,26].

2.3.4 Statistical analysis.

For each fold and each helminth species, we computed the following metrics at the object level:

Precision (positive predictive value), defined as

Recall (sensitivity or true positive rate), defined as

Average precision at an intersection over union of 0.50 (AP50), defined as the area under the precision–recall curve at a 50% overlap threshold [27].

Due to the design of our dataset, which exclusively targeted smears confirmed positive for STH or SCH to maximise annotation of helminth eggs for initial model training, we lacked negative-control slides (i.e., slides confirmed free of helminth eggs) necessary to compute diagnostic specificity or negative predictive value. While object-level precision, recall and AP50 characterise the model’s per-egg detection performance, slide-level specificity, repeatability and reproducibility will be assessed in future work incorporating egg-free specimens [20].

3.2.1 Hardware usability.

Although users considered the equipment transport case bulky, it provided robust protection for the equipment during air and ground transport. Fieldwork in Uganda required long daily commutes over bumpy and challenging dirt roads to reach various schools, where the equipment was set up and studies were conducted. Users expressed concern about the conspicuous appearance of the case, noting it could attract undue attention during transit. Additionally, on first use, it was unclear to some users that magnets secured the slide in the scanner tray, leading to potential handling issues. The lack of handles on the scanners made lifting them out of the case difficult, hindering setup efficiency. Lastly, cable clutter from an external router and loose connections was reported as a usability challenge in field settings. Recommendations to address these issues include reducing the case’s weight and volume while maintaining protection, adapting its outward appearance to resemble a standard suitcase to lower perceived value, redesigning the tray for greater tolerance of standard slides, adding handles to improve scanner portability, and integrating the router, battery backup, and Slide Manager into the scanner housing to minimise cable clutter.

3.2.2 Power and portability.

In laboratory settings, the UPS could sustain operations during power interruptions for up to 2.5 hours. For continuous use over an entire working day (8 hours) at field sites lacking mains power, portable diesel generators were predominately used. However, voltage fluctuations from these generators reduced the reliability and lifespan of the UPS in off-grid environments. Lithium power bank batteries were later found to be a more affordable alternative, offering longer operation times (up to 3.5 hours on a 20,000 mAh capacity) with significantly smaller volume and weight, and the ability to be charged simultaneously from common power sources including USB-C chargers or solar panels. Nevertheless, these batteries presented restrictions during air transport for field trials, as they were only permitted in carry-on luggage and not in checked baggage. Recommendations to address these challenges include incorporating voltage stabilisers or more robust power regulation to mitigate fluctuations, standardising around power banks for power input, clarifying air transport guidance, enabling hot swapping of power banks, and testing solar recharging options.

3.2.3 Scanning performance.

Users found the equipment setup straightforward, enabling them to assemble and initiate slide scanning in under 30 minutes with minimal guidance, training, or instructions. However, initial scan times of 25–35 minutes per slide were deemed too slow for efficient fieldwork, although later improvements reduced this to an average scan time of 12.5 (±6 SD) minutes per slide, with variability due to sample preparation and sample area with 622 (±102 SD) fields of view images per scan. While feedback from both countries emphasised a desire for faster scan times, stakeholders also valued the potential to parallelise scanning and automate data reporting and quality checks, which would free operators to focus on other tasks and improve overall cost efficiency. Additionally, hardware issues were noted, including one scanner becoming inoperable due to failed glue on the magnetic pickup for slides, preventing scanning operation. Recommendations include optimising the scanning algorithm and optics to further reduce scan time, as well as providing maintenance training, repair manuals, and local repair kits to address hardware failures. Other challenges observed during fieldwork, though not directly in Table 2, included susceptibility to insects entering WSI scanners during nighttime operations due to attraction to the bright light source, highlighting the need for an automatic closing flap on the slide loading port, and failures of Ethernet cables due to repeated setup and pack-down in outdoor environments, which were later mitigated by a software update enabling wireless image transmission.

3.2.4 Software and interface.

Users initially performed manual annotations of helminth eggs using the EggInspector application on a laptop, a laborious but essential task for developing the database for AI model training. This process was better suited to a laboratory or office environments and was most effective with a mouse due to the precision required to draw bounding boxes around objects; otherwise, users from both sites found the application user-friendly and easy to use. Once the AI model was trained, reviewing AI-detected objects on the laptop became more streamlined, focusing only on the egg-like objects. However, challenges included the lack of a high-contrast mode on the screen, which was problematic when operating in direct sunlight. Additionally, while the touch screens were capable of operation with disposable gloves, switching between tasks such as sample preparation, slide loading, and egg verification required careful consideration to glove disposal and equipment sterilisation. Users noted that verification work could be deferred to a later stage (e.g., after returning from the field) since images and data were preserved. Although laptop battery life was sufficient to sustain a full day of work, verifying eggs in field scenarios was considered potentially more suitable on a mobile device.

3.2.5 Training and support.

Users emphasised the value of practical training over written instructions, finding hands-on training sessions far more effective than paper manuals for mastering the AI-DP platform’s operation. While remote support was useful for troubleshooting, field maintenance remained challenging due to limited on-site resources and technical expertise. Recommendations include expanding training programs with local biomedical support materials and instructional videos, as well as developing a printable or online troubleshooting manual to assist with common issues. Additionally, options for engaging local support technicians should be explored to enhance field support capabilities.