Validation of Digital Slide Scanning and a Convolutional Neural Network for the Detection of Intestinal Parasites in Human Stool Samples

<b>Background</b>: Recent studies have shown that digital microscopy (DM) combined with a convolutional neural network (CNN) model is a valid approach for detecting intestinal protozoa and helminth ova or larvae in both trichrome-stained and wet-mount stool preparations. This study evaluated the diagnostic performance of a DM/CNN workflow for routine detection of intestinal parasites in a clinical microbiology laboratory. <b>Methods</b>: A clinical validation was conducted using the Grundium Ocus 40 scanner combined with the Techcyte Human Fecal Wet Mount (HFW) algorithm. The system was evaluated on (a) 135 reference samples and (b) 208 routine clinical samples submitted for intestinal parasite testing. Analytical sensitivity, precision, and limit of detection (LOD) were assessed. <b>Results</b>: For reference samples, the DM/CNN workflow achieved a positive slide-level agreement of 97.6% (95% CI: 94.4-100%), following a confidence threshold adjustment for <i>Schistosoma mansoni</i>, and a negative agreement of 96.0% (95% CI: 86.6-98.9%) compared with light microscopy (LM). Dilution series with reference samples revealed slightly lower analytical sensitivity of the DM/CNN at higher dilutions. Both intra- and inter-run precision studies demonstrated high reproducibility and stability. In prospective testing on 208 routine samples, overall agreement between DM/CNN and LM was 98.1% (95% CI: 95.2-99.2%) with a Cohen's Kappa coefficient of <i>κ</i> = 0.915. Minor discrepancies involved <i>Blastocystis</i> spp., with DM/CNN showing slightly higher sensitivity. <b>Conclusions</b>: For the first time, we show that the combination of the Grundium Ocus 40 scanner and the Techcyte HFW algorithm provides a reliable, low-throughput screening solution that can effectively assist diagnostic technicians by pre-classifying putative parasitic structures for targeted expert review. Despite its lower throughput, the system substantially reduces the manual review process and simplifies the parasitological workflow. Implementation in a clinical microbiology laboratory requires extensive site-specific validation to account for differences in sample processing and imaging conditions. Moreover, optimization of confidence thresholds for specific classifiers is essential to ensure consistent analytical performance across different laboratory settings.