Automatic 3D cell segmentation of fruit parenchyma tissue from X-ray micro CT images using deep learning

Abstract

Background: High quality 3D information of the microscopic plant tissue morphology-the spatial organization of cells and intercellular spaces in tissues-helps in understanding physiological processes in a wide variety of plants and tissues. X-ray micro-CT is a valuable tool that is becoming increasingly available in plant research to obtain 3D microstructural information of the intercellular pore space and individual pore sizes and shapes of tissues. However, individual cell morphology is difficult to retrieve from micro-CT as cells cannot be segmented properly due to negligible density differences at cell-to-cell interfaces. To address this, deep learning-based models were trained and tested to segment individual cells using X-ray micro-CT images of parenchyma tissue samples from apple and pear fruit with different cell and porosity characteristics.

Results: The best segmentation model achieved an Aggregated Jaccard Index (AJI) of 0.86 and 0.73 for apple and pear tissue, respectively, which is an improvement over the current benchmark method that achieved AJIs of 0.73 and 0.67. Furthermore, the neural network was able to detect other plant tissue structures such as vascular bundles and stone cell clusters (brachysclereids), of which the latter were shown to strongly influence the spatial organization of pear cells. Based on the AJIs, apple tissue was found to be easier to segment, as the porosity and specific surface area of the pore space are higher and lower, respectively, compared to pear tissue. Moreover, samples with lower pore network connectivity, proved very difficult to segment.

Conclusions: The proposed method can be used to automatically quantify 3D cell morphology of plant tissue from micro-CT instead of opting for laborious manual annotations or less accurate segmentation approaches. In case fruit tissue porosity or pore network connectivity is too low or the specific surface area of the pore space too high, native X-ray micro-CT is unable to provide proper marker points of cell outlines, and one should rely on more elaborate contrast-enhancing scan protocols.

Keywords: Artificial intelligence; Contrast-enhanced imaging; Fruit physiology; Image processing; Instance segmentation; Plant microstructure; X-ray micro-computed tomography.

Fig. 1

Stone cell clusters in pear tissue. (Top) 2D slices of contrast-enhanced micro-CT images of pear tissue samples of the inner cortex with stone cell clusters indicated with arrows. (Bottom) 3D visualization of the stone cell clusters

Fig. 2

Individual cells of (top) pear and (bottom) apple cultivars at different cortex positions with labels shown in colour scale for the cell volume

Fig. 3

Probability density plots of the cell equivalent spherical diameter (top) and sphericity (bottom) of the pear (left) and apple (right) cultivars at different cortex positions with the dot and line representing the mean and standard deviation

Fig. 4

Individual pores of pear (top) and apple (bottom) cultivars at different cortex positions with labels shown in colour scale for the pore mean radius

Fig. 5

PLS-DA score plot with the scores of tissue samples used in train (•, n = 36) and test (x, n = 18) set in colour code according to the pome fruit cultivar with 95% confidence ellipses

Fig. 6

Cell segmentation results for a tissue sample of A ‘Jonagold’ and B ‘Conference’ in the test set with labels shown in colour scale for cell volume. From left to right: input volume of 667 × 667 × 667 voxels in grayscale values, ground-truth segmentation as collected with the semi-automated cell segmentation protocol, segmentation results of deep learning-based models trained with slice spacing of 8, 512 and segmentation result of the benchmark method using the watershed algorithm. AJI Aggregated Jaccard Index

Fig. 7

2D slice of cell segmentation results for a ‘Celina’ tissue sample with vascular tissue and a stone cell in the test set with labels shown in colour scale for cell volume. From left to right: 2D slice of the input volume of 667 × 667 × 667 voxels in grayscale values, ground-truth segmentation as collected with the semi-automated cell segmentation protocol, segmentation results of deep learning-based models trained by slice spacing 8, 512 and segmentation result of the benchmark method using the watershed algorithm. AJI Aggregated Jaccard Index

Fig. 8

Aggregated Jaccard Index of the (left) apple and (right) pear tissue samples cell segmentations of the test set by the deep learning model trained on pome fruit data by slice spacing 8 for the different cultivars and cortex positions

Fig. 9

XY scores plot of the first latent variable in the PLS model to predict the segmentation results with colour code according to the pome fruit cultivar

Fig. 10

Regression coefficients of the predictor variables for the first latent variable in the PLS model, shown in colour scale, to predict the segmentation results, expressed by the Aggregated Jaccard Index. SSA specific surface area

Fig. 11

A Excision of pome fruit parenchyma tissue samples at inner, middle and outer positions on the fruit equator followed by B conventional X-ray micro-CT imaging. C Incubation of the tissue sample in a 10% (w/v) cesium iodide solution D followed by contrast-enhanced X-ray micro-CT imaging

Fig. 12

Data split into training, validation and test set