3084 – A MACHINE-LEARNING APPROACH TO SCORING ERYTHROID COLONIES IN SEMI-SOLID MEDIUM

Abstract

The ability of erythroid progenitors to form colonies in semi-solid medium is the ‘gold-standard' approach to establishing their identity. However, colony formation assays are laborious, requiring multiple replicates and manual scoring by skilled observers. These drawbacks limit the scale of their use in the laboratory, in fundamental research and in the search for novel erythroid stimulating agents. Here we present a novel tool, ‘c-count', which combines high-resolution automated image acquisition with a convolutional neural network (CNN) machine-learning model to accurately count colony-forming-unit-erythroid (CFU-e) colonies. We performed colony assays in a methylcellulose medium supplemented with erythropoietin (Epo) and iron-saturated transferrin, in 35 mm plates. Colonies were stained for expression of hemoglobin with diaminobenzidine. Images were acquired using the Zeiss Axio Observer inverted microscope. Sixteen contiguous fields were scanned and stitched together in each of 4 identically placed regions totaling 1/16th of each plate's area. To generate data to train and test the CNN model, a Laplacian of Gaussian algorithm was used to identify ‘blobs' within the images, which were then labeled by 5 independent observers as either true or false CFU-e. The training data contained 1062 blobs with 420 labeled as CFU-e, while the validation data contained 188 blobs. The F1 score among the 5 independent observers on the same test data ranged from 0.84 to 0.89 (mean=0.87). The F1 score of the c-count classification against the validation data was 0.87. In an additional test experiment, the correlations between CFU-e counts generated by c-count and two independent observers were 0.96 and 0.97. We are currently extending evaluation to determine the model's performance under a variety of experimental conditions including tissue origin of CFU-e and culture conditions. We conclude that c-count is a promising tool for scoring CFU-e with the potential to speed up the study of erythroid progenitor biology.