Comparison of object detection networks for cellular detection. Performance in terms of AP of Faster-RCNN (A), YOLOv2 (B), YOLOv3 (C), and RetinaNet (D) when trained on individual datasets without (TR1) and with vertically flipped data augmentation (TR2) and when trained using multiple datasets without (TR3) and with (TR4) vertically flipped data augmentation (n = 3, AP ± SD). Erythroblast DAPI cells (blue), Neuroblastoma phalloidin dataset (magenta), fibroblast nucleopore dataset (red), eukaryote DAPI dataset (orange), C127 DAPI dataset (green), and HEK peroxisome dataset (black). AP performance across all datasets for Faster-RCNN (E), YOLOv2 (F), YOLOv3 (G), and RetinaNet (H) for TR1–TR4. Additional vertical flipping of data (TR2) and joint training with multiple classes statistically boosts AP when using the Faster-RCNN networks, but not the other networks. Friedman’s test was applied using Dunn’s multiple comparison test to compare TR2–TR4 to T1 (n = 6, AP ± SD; *, P < 0.05; ***, P < 0.005).