During the last decade, non-conventional imaging techniques such as chlorophyll fluorescence imaging were used for the study of the interactions between plant and pathogens . Indeed, with chlorophyll fluorescence imaging, contrasts are enhanced compared to conventional color images, and depict more accurately the physiology of plant tissues [14, 15]. Chlorophyll fluorescence imaging provides images that map on leaves the variations of single parameters associated to photosynthesis. Among the various fluorescence parameters, we monitored variations in the maximum quantum yield of photosystem II photochemistry (Fv/Fm). We used chlorophyll fluorescence imaging to map on bean (P. vulgaris) leaflets the areas altered by Xff CFBP4834-R. It must be pointed out that most stresses that decrease leaf health will affect photosynthesis. Therefore we first checked that visible symptoms of CBB on bean leaflets co-localized with decreased values of the Fv/Fm parameter. Other parameters such as Fv/F0 or F0/Fm were shown to yield a high contrast between healthy tissues and tissues affected with various pathogens [20, 45, 46]. However, we did not treat these parameters in this study as they do not have a clear physiological significance .
Images based on the single Fv/Fm parameter are easier to segment by thresholding approaches than conventional color images, thereby easing the image analysis process. In such a context, we developed an automated thresholding procedure to select pixels corresponding to symptoms. The respective amounts of pixels corresponding to diseased or healthy areas can then be quantified to assess the disease severity on inoculated plants. Even though the decrease in Fv/Fm values due to the pathogen attack is now well documented [32, 33], only few studies developed approaches for the quantification of the diseased area on leaves. Most of the studies using the Fv/Fm parameter in plant pathology are based on the mean Fv/Fm value over the whole image [23, 33, 35–37, 47]. The mean Fv/Fm value may qualitatively discriminate between healthy and diseased leaves, but does not quantify the amount of diseased tissues [23, 33, 35–37, 47]. Only few studies attempted to analyze the pixel-wise Fv/Fm-distribution to discriminate between healthy and diseased organs using a threshold of Fv/Fm. For example, in the case of Fusarium culmorum, wheat ears were considered as infected when pixels displaying a Fv/Fm value lower than 0.3 could be observed in the image .
In the present study, we investigated thresholding approaches for the quantification of the diseased area on infected leaves. At first, trained raters compared a subset of visible images and Fv/Fm images of bean leaves of cv. Flavert inoculated with Xff CFBP4834-R. We could associate non-overlapping clusters of Fv/Fm values to each stage of the symptom development (necrotic, wilted, impacted and healthy tissues) caused by Xff CFBP4834-R on P. vulgaris cv. Flavert. Subsequently Fv/Fm thresholds could be defined to discriminate between the various stages of the symptom development. Counting the pixels associated to each Fv/Fm cluster enables the quantification of the leaf area corresponding to each stage of the symptom development on cv. Flavert. Defining non-overlapping clusters of Fv/Fm values to segment symptomatic areas can also be performed on a broad range of plant species to quantify areas affected by biotic or abiotic stresses. For example, on grapevine affected by lime-induced iron chlorosis, chlorotic areas displayed lower Fv/Fm values than healthy tissues .
Interestingly, lower Fv/Fm values may be observed on tissues located in the margin of symptomatic areas, but that do not display any visible symptoms. These areas evolve into symptoms over time. Therefore, as previously observed on Arabidopsis thaliana or Nicotiana benthamiana inoculated with Pseudomonas syringae, pre-symptomatic areas may also be phenotyped using Fv/Fm[33, 36]. Fv/Fm values in these tissues may not differ from those observed in chlorotic tissue and both chlorotic and pre-symptomatic tissues were grouped into impacted tissues in the present study. The decrease of the Fv/Fm values in pre-symptomatic areas is not fully understood. Indeed, neither these areas are yet colonized by bacteria, nor can be observed increased levels of ammonia or a restricting water movement .
However, non-overlapping clusters matching the various stages of the symptom development should be defined by trained raters in each pathosystem studied. Indeed, Fv/Fm clusters defined on Flavert do not match the visual observation on other cultivars. For example, the cluster corresponding to visually chlorotic tissues on cv. Michelet overlaps with that corresponding to necrosis on cv. Flavert (Additional file 1: Figure S1). Therefore fixed thresholds defined on a correspondence with visual observations by trained raters are valuable only within a single cultivar, and cannot be extrapolated to other cultivars. Using an expert-based thresholding approach on other cultivars needs a calibration step on a training set of pictures. Such a need is a limitation for this thresholding approach in the perspective of high throughput phenotyping, as visual assessment is time consuming.
Therefore, we aimed at defining Fv/Fm thresholds that could be extrapolated to any cultivar or plant species. As a decrease of Fv/Fm depicts the alteration of PSII, Fv/Fm threshold can be defined independently of visual observations. Clusters of Fv/Fm would depict objective stages of alteration of plant tissues. Moreover, the Fv/Fm parameter may be impacted by the physiological status of plants [49, 50] or abiotic stresses [16, 34, 51]. Fixed thresholds may therefore bias the quantification of the diseased area on leaves. Defining the thresholds on control plants for each experimental round helps avoid such a bias.
To avoid the use of fixed thresholds, we normalized our segmentation on mock-inoculated plants. We defined the threshold as the Fv/Fm value under which a healthy pixel only has a probability of 0.002 to be misclassified. Such a thresholding does not allow the discrimination of various stages of the symptom development. Therefore, within the diseased area, the pixel-wise Fv/Fm-distribution was modeled as a mixture of predicted Gaussian distributions. Such a modeling is largely used for image analysis in medical sciences [52, 53]. As well, in plant sciences such a modeling was recently applied to the automated recognition of individual Arabidopsis rosettes, in order to monitor independently the growth of each plant in the image . In the present study, the clustering of pixels according to Gaussian distributions aims at describing the various stages of alteration of plant tissues without any a priori based on visual observations. For each Fv/Fm image, a mixture of Gaussian is fitted independently, and such an approach does not need any calibration set of images. Based on these Gaussian mixture distributions, we could define non-overlapping clusters of pixels displaying similar Fv/Fm values, corresponding to the various stages of alteration of plant tissues.
Pathogen attack may also result in dwarfing or shrinking of leaves. Such a phenotype is rarely quantified [9, 12], but the use of non-destructive image analysis approaches may help solve such a caveat. In the present study, we monitored the size of leaflets over time. We considered the maximum size of each leaflet as a reference and the size decrease compared to this reference was considered as shrinking. Such a phenotyping is difficult to assess by visual observation only. A similar approach was used to analyze the leaf area impacted by herbivory . Other approaches were proposed to evaluate the leaf deformation, for example using a sphericity index [12, 55]. However, using such an index does not allow the quantification of the leaf area impacted by the shrinking.
To test the applicability of our segmentation approach for the evaluation of plant resistance, we quantified the symptoms caused by Xff CFBP4834-R on five commercial cultivars of bean (cvs. Flavert, Michelet, Pike, Wonder and Caprice). When looking at the total amounts of symptoms, the cv. Flavert appeared to be the most sensitive to Xff CFBP4834-R. The cvs. Caprice and Wonder exhibit few symptoms and can be considered as tolerant. The cvs. Michelet and Pike are impacted to an intermediary extent. On top of the amount of symptomatic tissues, selection for resistance may also focus on the stage of development of the symptom. Indeed, it may be of interest for breeders to notice that cvs. Caprice and Wonder exhibit different symptoms topologies, even though they displayed similar total amounts of symptomatic tissues.
Finally, in order to select for quantitative plant resistance to pathogens, high throughput procedures aiming at precisely quantifying disease severity need to be developed. Robotic imaging procedures can increase the number of images taken  but few automatic analysis of chlorophyll fluorescence images procedures are available. The procedure presented in this study was automated under R and the R script is available at http://lisa.univ-angers.fr/PHENOTIC/telechargements.html. Running our procedure on the 1080 images of our dataset, two minutes only are needed for the Expert- and Probability-based thresholding analyses. The use of MCLUST  to discriminate various stages of alteration of plant tissues increases up to one hour the calculation time, which remains much faster than rating disease severity by visual observations.