Figure 3.

Deep learning workflow for Pinus radiata somatic embryo segmentation. Images are captured under a high-resolution microscope before being manually annotated to train and evaluate the two neural networks. For Mask R-CNN instance segmentation, cotyledon instance predictions are combined to derive a segmentation mask for direct comparison of pixel-wise metrics with ResNet semantic segmentation. Additionally, individual instances detected in boxes allow for cotyledon counts to be derived and a range of performance metrics are evaluated.