Figure 3: Overview of results.
Results for Tak Fa 1, Tak Fa 2 and Tak Fa 3 are shown in rows 1), 2) and
3). Column a) shows mean root hair length by treatment; column b) shows
distribution of root hair length by replicate and treatment; column c)
shows root hair density by treatment; columns d) and e) show
correlations between automatic and manual measurements of root hair
length and density, respectively.
CONCLUSIONS
We demonstrated an algorithm to resolve and extract occluded hair-like
structures on the example of root hairs captured in 2D microscopy
images. The high correspondence of our measurements to the validation
data shows that we are able to accurately determine the length and
density of root hairs in microscopy images. We observed that
correlations between automatic and manual measurement are lower in the
demonstration dataset with maize images than in our validation dataset.
Higher correlations in our validation set suggests that improving our
classification step could result in overall enhanced performance and
that manual measurements of root hairs are potentially not
representative. We further observed that manually measured root hair
length resulted in longer root hairs compared to automatically measured
root hairs, which suggests that selecting root hairs manually is biased
towards longer root hairs. We tested for difference between groups in
root hair density and mean root hair length per replicate (i.e. per
plant). While we could not determine significant differences between
genotypes and treatments in this dataset, we believe that our method
paves the way towards better identifying the genetic control of root
hair traits and an improvement in breeding programs for these traits. We
anticipate that incorporation of the measured trait distributions
(Figure 3, column b) into future statistical analysis will provide more
insight into root hair response to abiotic stresses. Further development
of our algorithm will allow to resolve intersections in branching
architectures with several orders of branching hierarchy to accurately
determine traits at larger organismal levels in the future.