Quantifying Root Colonization in Arbuscular Mycorrhizas by Image Segmentation and Machine Learning


 MotivationArbuscular mycorrhizas are the most widespread plant symbioses and involve the majority of crop plants. The beneficial interaction between plant roots and a group of soil fungi (Glomeromycotina) grants the green host a preferential access to soil mineral nutrients and water, supporting plant health, biomass production and resistance to both abiotic and biotic stresses. The nutritional exchanges at the core of this symbiosis take place inside the living root cells, which are diffusely colonized by specialized fungal structures called arbuscules. For this reason, the vast majority of studies investigating arbuscular mycorrhizas and their applications in agriculture require a precise quantification of the intensity of root colonization. To this aim, several manual methods have been used for decades to estimate the extension of intraradical fungal structures, mostly based on optical microscopy observations and individual assessment of fungal abundance in the root tissues. ResultsHere we propose a novel semi-automated approach to quantify AM colonization based on digital image analysis and compare two methods based on image thresholding and machine learning. Our results indicate in machine learning a very promising tool for accelerating, simplifying and standardizing this critical type of analysis, with a direct potential interest for applicative and basic research.Contactivan.sciascia@unito.it; andrea.genre@unito.it


36
Arburscural mycorrhizas (AM) are widespread plant endosymbioses that develop between 37 Glomeromycotina fungi and the roots of the majority of plants species, including most crops. The  In an attempt to improve speed, repeatability and reliability of this type of analysis, we developed a 59 semi-automated approach based on digital imaging and different types of post processing. Firstly, 60 an automatic method was developed to discriminate between mycorrhizal and non-mycorrhizal root 61 images. Secondly, we designed a semi-automated algorithm to generate quantitative indexes of root 62 colonization deriving from either image thresholding using ImageJ, or based on machine learning 63 analyses 8, 9 using the commercial software Zeiss Intellesis 10 . Our results indicate machine learning 64 as the most effective approach, with interesting applicative perspectives as an alternative to manual 65 quantification.

97
-a second dataset consisting of 180 images was used for both the thresholding (using ImageJ) 98 and machine learning (using Intellesis) approaches. All images had previously been classified 99 manually into 6 classes of 30 images, ranging from 1 (non mycorrhizal) to 6 (maximum root 100 colonization), adapting to the same standard protocol 6 .

101
The digital images in grayscale and the relative thresholding are described in Supplementary   102 file 2. Binary segmentation analysis.

107
In the first phase of the analysis, images were loaded into ImageJ 11 and transformed into 8-bit 108 digital images before further processing. A simple segmentation technique based on pixel intensity 109 thresholding was then applied in order to distinguish darker areas (likely corresponding to stained 110 fungal structures) from lighter areas (not colonized tissues). In more detail, the pixel brightness   Thresholding analysis. 132 Also in this case digital images were loaded into ImageJ and transformed into 8-bit digital images.

133
Two ImageJ macros were designed to process the images (Supplementary file 3). The first image 134 segmentation macro was designed to discriminate the entire root section area from the globally 135 lighter background; in this case the pixel intensity threshold was set to 155.

136
The second macro produced image segmentation based on pixel intensity, with a threshold 137 empirically set at 65 (in the 0-255 range) to discriminate between darker pixels (corresponding to 138 fungal structures) and lighter (uncolonized) areas.

139
Both results are described in Figure 3.  The ratio between these two values was used to obtain the m index (for mycorrhization) for each 151 image in the dataset: This m-index was therefore used in the subsequent statistical analyses as an independent variable in 154 the predictive model for the quantification of root colonization (Table 1). In the Table 1a we show the descriptive statistics of the calculated m-index based on ImageJ macros 159 relative for each manual classified classes in the range 1 -6. Confidence intervals and qualitative 160 descriptions for each class are shown in Table 1b.  Indeed, a strong correlation was detected between the two variables, with a significant Pearson 168 coefficient at the 0.01 two tailed level with a value of 0.748. Furthermore, ANOVA analysis (Table   169 2) and post hoc test of multiple comparisons performed with Bonferroni correction (  In order to build a prediction model of the level of mycorrhization, we used regression analysis with 179 linear, quadratic or cubic models ( Figure 5 and Table 4). Our results show that the most satisfactory  The independent variable is thresholding_index.  image background (Figure 6).

211
Once the training was completed, the model was applied to the entire dataset of 180 images.

212
Intellesis outputs different quantitative parameters of the areas in square pixels as in the case of the 213 thresholding technique explained above.

214
Also in this case we built a relationship index between segmented areas.

215
Original digital image Segmentation In this case, the ml-index for each image in the dataset is: The ml-index was therefore used in the subsequent analyses for the construction of a prediction 220 model to evaluate the level of mycorrhization.

221
In the  And the box plot of the descriptive statistics is described in Figure 7.    In the Table 8 below, R squared for ml-models are resumed. The independent variable is spec_segm_mlearning_index. We summarize in Table 9 the results of the statistical analysis performed comparing thresholding 257 and machine learning training using the best predictive cubic models.  The machine learning method, based on the Zeiss Zen Intellesis application, resulted to be the most 307 efficient. In this case, individual fungal structures such as intracellular hyphae and arbuscules were 308 manually selected during the training phase to generate a model that the software then applied to the 309 whole data set. This method identified 6 classes of mycorrhization intensity, achieving the best 310 correlation (Pearson correlation coefficient 0.824) with manual analysis.

311
Importantly, the training process was relatively quick (it required 50 minutes overall) and resulted 312 to be effective even when using a limited number of images (10 images). Lastly, the use of 313 machine-learning identification of shapes and objects allowed a reliable discrimination between 314 equally stained structures such as extraradical hyphae and arbuscules, a capability that our image 315 thresholding methods could never achieve.  In conclusion, it should be stressed that this is one of the first attempt to use a machine learning  Brightness-based selection of pixels on the same image of a mycorrhizal root segment. Pixels are selected (in red) based on arbitrary brightness thresholds: 100, 137, 170 in a range from 0 (black) to 255 (white) using Fiji/ImageJ.

Figure 3
Image thresholding in ImageJ. The rst threshold set at 65 identi es the colonized area, while the second threshold at 155 outlines with a good approximation the total area of the root section. Curve t with linear, quadratic and cubic regression models t index