Point-of-care platform integrated with deep-learning, convolutional neural network algorithms effectively evaluates...

To perform a diagnostic assessment of a point-of-care veterinary multiuse platform integrated with a model comprised of deep-learning, convolutional neural network algorithms for evaluating canine/feline peripheral blood smears compared to board-certified clinical pathologists (CPs).

This study had a blinded, randomized, incomplete block design, and results were compared between CPs and algorithms. Blood smears from convenience samples from veterinary diagnostic reference laboratories from October to December 2021 were used. Study phase A comprised 2 parts: (1) object class identifier algorithm (leukocytes, platelets, polychromatophils, and nucleated erythrocytes) versus CP within the same field of view (FOV); and (2) monolayer detection algorithm plus object class identifier algorithm versus CPs with different FOVs. Study phase B comprised algorithms versus CP for platelet clump identification. Study phase C comprised algorithms versus CP for polychromatophil identification. Metrics including sensitivity, specificity, and agreement were used.

The sample size was 59 dogs and 60 cats in phase A, 92 dogs and 69 cats in phase B, and 47 dogs and 12 cats in phase C. For study phase A, part 1, the 5-part leukocyte differential count agreement was 96.6% for canine and 91.7% for feline blood smears, and for part 2, the agreement for estimated total leukocyte, platelet, polychromatophil, and nucleated erythrocyte counts ranged from 70% to 95% across species. In study phase B, the algorithm had 90% sensitivity and 88% specificity. The algorithm for polychromatophils had 100% agreement with CP results in phase C.

This platform achieved results comparable to those of CPs. Results are meant to complement automated CBC results.

Veterinarians may add this assessment as part of their standard in-clinic hematology analysis for patients.