Erythrogram directly from the microscope eyepiece: a feasibility study using artificial intelligence

BackgroundErythrocyte indices are essential for the diagnosis and monitoring of hematologic diseases, but their determination depends on automated hematology analyzers, which limits access in regions with limited laboratory infrastructure. Although artificial intelligence approaches have been proposed for hematologic analysis, they usually rely on slide scanners or digitization systems. To date, no validated approaches have been identified in the literature that estimate these indices directly from images obtained through the eyepiece of conventional optical microscopes. ObjectiveTo evaluate the feasibility of automated prediction of erythrocyte indices from blood smear images obtained directly through the eyepiece of conventional microscopes using convolutional neural networks. MethodsTwo hundred blood samples stained using the May-Grunwald-Giemsa method were analyzed and photographed using a standard optical microscope. Four architectures, DenseNet-121, EfficientNet-B0, ResNet-18, and ResNet-34, were evaluated at different resolutions using 10-fold K-Fold cross-validation. ResultsFor RBC, HGB, and HCT, ResNet-34 at a resolution of 1024x1024 pixels achieved superior performance, with R2 between 0.90 and 0.92, Pearson correlation r > 0.95, and mean absolute errors of 0.184 x106/{micro}L, 0.524 g/dL and 1.292%, respectively. For RDW-CV, DenseNet-121 achieved R2 = 0.49 and r = 0.71, reflecting the greater complexity of this parameter. Bland-Altman analysis confirmed adequate agreement, with biases close to zero and more than 94% of observations within the limits of agreement. ConclusionArtificial intelligence demonstrated excellent predictive performance in estimating the erythrocyte indices RBC, HGB, and HCT, with R2 > 0.90, from images obtained using a conventional microscope and accessible hardware. This approach has significant potential to democratize access to hematologic analysis in resource-limited settings, although multicenter validation and regulatory evaluation are required before clinical implementation.