Three dimensional reconstruction of plant roots via low energy x-ray computed tomography
Plant roots are vital organs for water and nutrient uptake. The structure and spatial distribution of plant roots in the soil affects a plant's physiological functions such as soil-based resource acquisition, yield and its ability to live under abiotic stress. Visualizing and quantifying roots' configuration below the ground can help in identifying the phenotypic traits responsible for a plant's physiological functions. Existing efforts have successfully employed X-ray computed tomography to visualize plant roots in three-dimensions and to quantify their complexity in a non-invasive and non-destructive manner. However, they used expensive and less accessible industrial or medical tomographic systems. This research uses an inexpensive, lab-built X-ray computed tomography (CT) system, operating at lower energy levels (30kV-40kV), to obtain two-dimensional projections of a plant root from different viewpoints. I propose image processing pipelines to segment roots and generate a three-dimensional model of the root system architecture from the two-dimensional projections. Observing that a Gaussian-shaped curve can approximate the cross-sectional intensity profle of a root segment, I propose a novel multi-scale matched filtering with a two-dimensional Gaussian kernel to enhance the root system. The filter assumes different orientations to highlight the root segments grown in different directions. The roots are isolated from the background by manual thresholding, followed by a mathematical morphological process to reduce spurious noise. The segmented images are filtered back projected to generate a three-dimensional model of the plant root system. The results from the research conducted show that the proposed method yields a structurally consistent three-dimensional model of the plant root image set obtained in the air, whereas alternate methods could not process the image set. For plant root images collected in the air, the three-dimensional model generated from the proposed matched-guided filtering and filtered back projection has a better contrast measure (0.0036) compared to the contrast measure (0.099) of the three-dimensional model created from raw images. For plant root images captured in the soil, proposed multiscale matched filtering resulted in better receiver operating characteristic curves than the raw images. Compared to Otsu's thresholding, multi-scale root enhancement and thresholding have reduced the average false positive rate from 0.344 to 0.042, and improved the average F1 score from 0.4 to 0.775. Experimental results show that the proposed root enhancement methods are robust to the number of orientational filters chosen, and are sensitive to the filter length selected. Small size filters are preferred, since increasing the filter length increases the number of false positives around root segments.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 License.