A generalised approach for high-throughput instance segmentation of stomata in microscope images

Abstract Background Stomata analysis using microscope imagery provides important insight into plant physiology, health and the surrounding environmental conditions. Plant scientists are now able to conduct automated high-throughput of stomata in data, however, existing detection methods sensitive appearance training images, thereby limiting general applicability. In addition, only generate bounding-boxes around detected stomata, which require users implement additional image processing steps study morphology. this paper, we develop a fully automated, robust algorithm can also identify individual boundaries regardless species, sample collection method, imaging technique magnification level. Results The proposed solution consists three stages. First, input is pre-processed remove any colour space biases occurring from different techniques. Then, Mask R-CNN applied estimate boundaries. feature pyramid network embedded utilised at scales. Finally, statistical filter implemented output reduce number false positive generated by network. was tested 16 datasets 12 sources, containing over 60,000 stomata. For first time domain, against 7 never seen show generalisability solution. indicated that approach detect with precision, recall, F-score 95.10%, 83.34%, 88.61%, respectively. A separate test conducted comparing estimated boundary values manually measured data showed method has an IoU score 0.70; 7% improvement bounding-box approach. Conclusions shows performance across multiple quality scale. This generalised allows whilst eliminating need re-label re-train for each new dataset. open-source code shared project be directly deployed Google Colab or other Tensorflow environment.