Fig. 3From: Deep learning identifies Acute Promyelocytic Leukemia in bone marrow smearsPerformance of convolutional neural nets for binary cell type classifications. Since end-to-end image-level classification did not show satisfactory results in preliminary testing, we used cell-level recognition with convolutional neural nets as a proxy. Relevant cell types and features for the distinction between non-APL AML, APL and healthy bone marrow, i. e. myeloblasts, promyelocytes, and Auer rods, were labeled manually and CNNs were trained. The performance of individual CNNs for the detection of myeloblasts (A), promyelocytes (B), and Auer rods (C) on the respective testing sets (that were rigorously withheld from training) is displayed as area under the receiver operating curve (AUROC) using three-fold cross-validation (cv 0, 1, 2 illustrated in light blue, orange, and green). std. dev. – standard deviation of the mean; TPR – true positive rate; FPR – false positive rateBack to article page