Stomata segmentation using deep learning

Resumen

Stomata are pores in the epidermal tissue of leaf plantsformed by specialised cells called guard cells, which regu-late the stomatal opening. Stomata facilitate gas exchange,being pivotal in the regulation of processes such as pho-tosynthesis and transpiration. The analysis of the numberand behaviour of stomata is a task carried out by study-ing microscopic images; and, nowadays, this task is mainlyconducted manually, or using programs that can count anddetermine the position of stomata but are not able to deter-mine their morphology. In this paper, we have conducteda study of 10 deep learning algorithms to segment stom-ata from several species. The model that achieves the bestDice score, with a value of 96.06%, is obtained with theDeepLabV3+ algorithm, whereas the model that providesthe best trade-off between inference time and Dice scorewas trained using the ContextNet architecture. This is afirst step towards improving the measurements provided bystomata analysis tools, that will in turn help plant biologiststo advance their understanding of dynamics in plants