Download Auto Root Tools For Pc

Computer-based phenotyping of plants has risen in importance in recent years. Whilst much software has been written to aid phenotyping using image analysis, to date the vast majority has been only semi-automatic. However, such interaction is not desirable in high throughput approaches. Here, we present a system designed to analyse plant images in a completely automated manner, allowing genuine high throughput measurement of root traits. To do this we introduce a new set of proxy traits.

We test the system on a new, automated image capture system, the Microphenotron, which is able to image many 1000s of roots/h. A simple experiment is presented, treating the plants with differing chemical conditions to produce different phenotypes. The automated imaging setup and the new software tool was used to measure proxy traits in each well. A correlation matrix was calculated across automated and manual measures, as a validation. Some particular proxy measures are very highly correlated with the manual measures (e.g. proxy length to manual length, r2 > 0.9). This suggests that while the automated measures are not directly equivalent to classic manual measures, they can be used to indicate phenotypic differences (hence the term, proxy). In addition, the raw discriminative power of the new proxy traits was examined. Principal component analysis was calculated across all proxy measures over two phenotypically-different groups of plants. Many of the proxy traits can be used to separate the data in the two conditions.

Download File 🔥 https://fancli.com/2y2Q1u 🔥

The new proxy traits proposed tend to correlate well with equivalent manual measures, where these exist. Additionally, the new measures display strong discriminative power. It is suggested that for particular phenotypic differences, different traits will be relevant, and not all will have meaningful manual equivalent measures. However, approaches such as PCA can be used to interrogate the resulting data to identify differences between datasets. Select images can then be carefully manually inspected if the nature of the precise differences is required. We suggest such flexible measurement approaches are necessary for fully automated, high throughput systems such as the Microphenotron.

The problems of shoot and root phenotyping present distinct challenges, as such many tools are developed with a focus on roots specifically. For truly automated phenotyping, bottom-up approaches are the most common. GiaRoots [9], EzRhizo [10], WinRhizo and RootReader2D are notable examples. These often apply a low level binary classification operation such as local or global thresholding, to determine which pixels correspond to root material, and which to background. The benefit of assigning labels to pixels in this way is that broad quantitative measures of root systems can be calculated quickly. Measurements such as inferred rootmass, width, tortuosity etc., are easily computed on binary images. However, it is often challenging to perform this thresholding in the presence of noisy images, and large amounts of processing can be required to clean up the signal, for example by removing anomalously-detected foreground pixels. In those cases where the removal process is not perfect, the resultant trait values become flawed. It is commonly held that this low-level error can be overcome by increasing the size of the dataset, something that can be done trivially in automated systems. This approach undoubtedly has merit, but the extent to which this is true in practice will often be a function of the input images, and of the traits being measured; care must be taken.

To adopt a fully automated phenotyping approach, any software must fulfil particular criteria. Once running it should not make any demands on the user; all images in the set must be processed in one go, as a batch. Previous software has taken this approach (e.g. [17]), requiring the user to provide some details to the software initially, but after that period, batch processing proceeds through an entire set of images. However, it is still beneficial to perform manual visual checking of the final results, to confirm whether the images have been successfully processed. At some point, though, this approach becomes infeasible. With the introduction of more capable robotics enabling very high throughput image capture, it becomes challenging to verify that all the results are satisfactory. We need to ask the software to either place a confidence in the measured results, or provide results which are inherently probabilistic. To say a root is 32.6 mm long requires much certainty on behalf of the software and developers, and the degree of this certainty is often not addressed in automated software. To say root A is longer than root B may be just as valuable, yet requires a looser set of processing requirements.

The software itself should run sufficiently fast so as to keep up with throughput of the image capture, or at least be able to batch process results offline and in time for the arrival of the next batch. With typical phenotyping studies requiring 1000s of images (e.g. [18]) the issue of processing speed is becoming increasingly important. This is the motivation behind the approach presented here. We propose a system designed explicitly to work with images generated in a high throughput manner using a robotic capture system. The software does not require pre-training per image set, as image capture settings are able to be kept consistent as part of the imaging setup. Processing requirements are sufficiently small that an image can be processed in a few seconds on a standard PC (i.e. faster than the rate of image capture). User interaction is not required during processing, and results are automatically generated.

where dijkstra(x, y) is the shortest distance to pixel x, y, from any start position (see Fig. 1). The value L is normalised using the maximum theoretical value of distance for an image of that size. In practice this can be calculated as the maximum distance returned by Dijkstra, averaged over a number of wells. Optionally, the function L can also be raised to a power, i.e. L n to decrease the distance at which the likelihood drops off from bright pixels. We have found this has minimal impact on results, but can be useful if plants are more established, and thus the maximum distance from root material is never large.

a Top-down image from the Microphenotron. b Side-facing image of the same plants in a. c L(x, y) visualised as a heat map for the set of eight wells in b. Brighter areas indicate regions the software considers more likely to be root material

The function L indicates a likelihood of there being a root at a given location. Note, we are not thresholding the image at this point, rather we wish to assign a confidence level to represent how certain we are that this pixel belongs to a root. Any subsequent trait derived from this data can maintain the idea of this confidence. Therefore, we avoid the problematic situation of having to determine a priori the exact existence or non-existence or root at each location in the image, and can instead make an educated guess.

Once image pixels are assigned a confidence level indicating root material presence, and a distance to the determined anchor point, a number of interesting, but non-traditional traits can be measured from the image. They are measured in the image space (i.e. in pixels) but as we will see this is not important and conversion to real-world units is not necessary. The measures currently implemented in AutoRoot are presented in Table 1.

These traits are calculated as a weighted function (e.g. a sum or average), where the contribution of each pixel to the final measurement is given by its likelihood. So rather than, for example, measuring the mass of the root system by summing over all thresholded foreground pixels, we sum the likelihoods of all pixels in a well. This means that all pixels are considered within each measured trait, but that those with high likelihood of being root material will make a significantly higher contribution. The orientation \(\theta\) at each pixel is calculated using a Sobel convolution:

In the following experiments we examine how useful the new metrics are as proxy measures for classically measured traits via a simple phenotyping experiment. We also consider the usefulness of using the new proxy measures directly as ways of discriminating different phenotypes. We have purposefully chosen clearly visible and distinct phenotypes in order to demonstrate both the process and descriptive power of the proxy traits. AutoRoot has been used successfully to recover more subtle phenotypic differences in an Arabidopsis root and shoot chemical screening experiment [20].

As well as being correlated with traditional traits, the new proxy measures are useful discriminative measures in their own right. To demonstrate this, we analysed the raw data produced by the software from the same growth experiment described above. All well plates in all images were analysed, fully automatically without user interaction. Scatterplots of pairs of proxy measures (commonly referred to as a pairs plot) were produced (Fig. 3). This figure gives a helpful indication of how useful the measures are in identifying phenotypic differences in our example dataset. As can be seen in Fig. 3, the two growth conditions are strongly separable in many of the newly proposed measures. Based on this, traditional unsupervised clustering approaches such as k-means would work effectively in separating phenotypic classes.

In this paper we present an alternative to traditional manual image measures, which we term proxy traits. They can be calculated over complex images where segmenting all parts of a root system is not possible, or not reliable. These proxy traits are well suited to fully automated analysis settings, especially as part of automated robotic-based systems. Computational performance exceeds image capture, so does not produce an additional bottleneck. ff782bc1db