A prospective study for the examination of peripheral blood smear samples in pediatric population using artificial intelligence.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
BACKGROUND/AIM: Peripheral blood smear (PBS) and bone marrow aspiration are gold standards of manual microscopy diagnostics for blood cell disorders. Nowadays, data-driven artificial intelligence (AI) techniques open new perspectives in digital hematology. This study proposes an AI learning technique for the classification of blood cells over PBS samples while increasing the sensitivity and specificity rates of the experts as a decision support system of a prediagnostic tool. MATERIALS AND METHODS: The methodology of this study comprises three steps for the creation of an effective learning technique for blood cell disorders. First is the digitization of PBS samples in 100x optical-digital magnification using Mantiscope which is a cloud-based slide scanner system. The second is collection of pediatric hematology experts' annotations and the last one is data augmentation to increase the data variation and size. The data consists of 372 individuals, an approximate number of 12,000 annotated images with 500,000 blood cell objects. A subjective test is also performed to observe the interobserver variability. RESULTS: We measured sensitivity and specificity for 28 cell types for the resulting decision support system. We obtained sensitivity 98% for myeloblast, 94% for basophil and 90% for lymphoblast, specificity 99% for basophil, eosinophil, monocyte, hypersegmented neutrophil, band neutrophil and reactive neutrophil in leukocyte subtypes. When erythrocyte measurements were evaluated, it was found that the sensitivity was 93% for normoblast, 81% for target cell and pencil cell, 80% for sickle cell, specificity was 99% for normoblast, pencil cell, echinocyte, and sickle cell. CONCLUSION: It is observed that sensitivity and specificity greater than 90% can be obtained for some specific cell types with this clinical study. It is seen that data augmentation increases the effectiveness of the learning method in terms of leukocytes by improving the measurement metrics. This could be a valuable technique to evaluate acute leukemias and hemolytic disorders.