Figure 1.

Schematic of the segmentation and analysis pipeline. Reconstructed microCT scans are manually thresholded to find the best value to segment the airspace of the leaf (as in Théroux‐Rancourt et al., 2017). Using this binary stack, a local thickness stack is created, which identifies for each pixel the diameter of the largest sphere contained in that area (lighter pixel values mean larger diameters). These input stacks are used to generate the feature layers arrays needed, along with the hand‐labeled slices, for the random forest classification model training. With the trained model, the complete stack of images is predicted, and from this predicted stack the image is post‐processed to remove false classifications, and leaf traits are analyzed. Note that all images are from the same position within the stack (i.e., same slice) except for the segmented image—using images from the same slice used for hand labeling provides identical segmentation.