Detection of protozoan and helminth parasites in concentrated wet mounts of stool using a deep convolutional neural...

Comprehensive diagnosis of gastrointestinal parasites is largely reliant on traditional stool microscopy, despite gains in molecular diagnostics. Wet-mount examinations remain a significant challenge for traditional microscopy, digital microscopy, and artifical intelligence (AI). We developed and validated a deep convolutional neural network (CNN) model that provides highly sensitive detection and presumptive classification of enteric parasites. Twenty-seven different parasites were trained on a CNN model using a wide diversity of 4,049 unique parasite-positive specimens (determined by traditional microscopy) collected in the USA, Europe, Africa, and Asia. Model validation was performed with a unique holdout set. In clinical validation, AI correctly detected 250/265 positive specimens (94.3% agreement) and 94/100 negative specimens (94.0%) before discrepant resolution. AI also detected 169 additional organisms from the validation specimens that were not previously identified. These additional detections underwent further discrepant analysis to adjudicate the results by scan review and microscopy. After resolution and inclusion of newly defined true positives and false positives, the positive agreement was 472/477 (98.6%). Negative agreement was variable by organism, ranging from 91.8% to 100%. A relative limit of detection study was performed comparing AI to three technologists of varying experience using serial dilutions of specimens containing Entamoeba, Ascaris, Trichuris, and hookworm. AI consistently detected more organisms and at lower dilutions of parasites than humans, regardless of the technologist's experience. The use of AI for wet-mount analysis is highly sensitive and detects considerably more organisms than traditional microscopy alone. The use of AI simplifies the parasitology workflow and reduces the reliance on traditional microscopy.IMPORTANCEGastrointestinal parasite ova and parasite (O&P) detection from stools is a manual, labor-intensive method requiring highly trained personnel. This testing has been largely unchanged in 100 years, with the exception of minor improvements in processing and fixation techniques. O&Ps are performed worldwide on millions of stool specimens a year, making any improvements in the process highly impactful. Digital slide imaging and artificial intelligence were recently established tools by our laboratory for improving permanent trichrome stain interpretation. This work builds on that breakthrough and describes the first comprehensive wet-mount AI model development and validation. Improved diagnostic yield, analytical sensitivity, and precision were demonstrated in this work through full clinical laboratory validation studies, including a specimen collection sourced from four continents and a diversity of fixatives and preparation techniques. This work represents the completion of a groundbreaking effort to bring parasite screening into the technological age.