Here, we demonstrate how X-ray micro-computed tomography and a random forest machine learning algorithm can be coupled to quantify plant starch content in vivo, repeatedly and non-destructively over time in three grapevine species.
We began this project after observing apparently depleted regions in the stems of parenchymal tissue of grapevine plants. Here's a panel of images demonstrating what we saw that piqued our curiosity.
We suspected that these depleted regions correspond with starch depletion in ray and axial parenchyma (RAP) cells. If so, this would be very exciting, as no techniques currently exist for quantifying and spatially mapping plant starch reserves in-vivo.
To help us quantify the presence/absence of these full versus empty RAP regions, we developed the following machine learning framework:
First, X-ray microCT images are collected, resulting in 32-bit images of the stem cross section for each plant. Visually empty/full parenchymal regions are manually labeled as full or empty of starch (see image below). Manually labeled images are split equally into test and training image datasets. MicroCT images are preprocessed (i.e. cropped, denoised, and contrast stretched) to normalize images across plant samples and to facilitate learning by the training algorithm. Feature layers are generated by convolving the preprocessed images with various types of kernels (e.g. Gaussian, variance, lines, and patches) that correspond with spatial patterns in X-ray absorption of starch, parenchymal cells, and cell wall tissue. A random forest algorithm is used to train a model to predict the labeled training images based on available feature layers. The trained model then predicts empty/full RAP regions in test images that are not used for training and the model’s performance is evaluated.
Here's what the final predictions look like for all the RAP cells:
These predictions correlate well with enzymatically measured starch concentration (R^2 = 0.95).
Earles, J.M., Knipfer, T.K., Tixier, A., Orozco, J., Reyes, C., Zwieniecki, M.A., Brodersen, C.R., and McElrone, A.J. (accepted). In-vivo quantification of starch reserves in plants using X-ray microCT imaging and machine learning. *Authors contributed equally
Note: The image datasets used in Earles et al. (accepted) are several Gb, so we aren't hosting them on GitHub. Please send us a message if you would like a direct link to download them.