On-field optical imaging data for the pre-identification and estimation of leaf deformities

Labeled temperate hardwood tree stomatal image datasets from seven taxa of Populus and 17 hardwood species

Introduction

Wood is a complex biological structure providing physical support and serving the needs of a living plant. Its degradation by stresses or pathogens exposes the plant to huge physiological and structural changes, but the consequences might not be immediately detectable from the outside. Trunk diseases can spread internally and silently, degrading woody tissues, then erratically leading to external symptoms, production losses and, ultimately, the plant's death. While accurate detection and monitoring of wood degradation require structural and functional characterization of internal tissues, collecting such data on entire plants—nondestructively and in-vivo—is extremely challenging. As a result, the study and management of wood diseases are almost impossible in-situ. These diseases generate enormous losses on a global scale, including for sectors such as the wine industry.

Grapevine trunk diseases (GTDs) are a major cause of grapevine (Vitis vinifera L.) decline worldwide¹. They are mostly undetectable until advanced stages are reached, and the European Union has banned the only effective treatment, i.e., an arsenic-based pesticide. Therefore, vineyard sustainability is jeopardized, with yearly losses of up to several billion dollars². Detecting and monitoring GTDs is extremely difficult: fungal pathogens insidiously colonize trunks, leaving different types of irreversibly decayed tissues¹. The predominant GTD, Esca dieback, erratically induces tiger stripe-like foliar symptoms, but their origin remains poorly understood³, and their sole observation is not indicative of the vines' sanitary status^4,5,6,7. Quantifying degraded tissues within living vines could help determine the plant’s condition and predict disease evolution. However, classical techniques⁸ require sacrificing the plant and often yield limited information. Reaching a reliable diagnosis is thus impossible in

living plants.

Monitoring wood degradation using non-destructive imaging techniques has primarily been tested on detached organs, blocks, or planks, and using a single technique: Magnetic Resonance Imaging (MRI)^9,10 or X-ray computed tomography (CT)^11,12,13,14. In grapevine, CT scanning allowed the visualization of the graft union¹⁵, the xylem refilling¹⁶, tyloses-occluded vessels¹⁷, and the quantifying of starch in stems¹⁸. Synchrotron X-ray CT has been applied to leaves to investigate the origin of foliar symptoms related to trunk diseases, suggesting that symptoms might be elicited from the trunk³. X-ray CT and MRI were successfully combined to collect anatomical and functional information to investigate stem flow in-vivo¹⁹. These techniques were recently tested for trunk disease detection²⁰ but were applied separately and on different wood samples, preventing the possibility of combining modalities and thus limiting their effectiveness, and failing to distinguish healthy and defective tissues using MRI.

Imaging-based plant studies are usually performed at a microscopic scale and on a specific detached tissue or organ but rarely on a whole plant. The field of investigation is limited by difficulties in adapting imaging devices to plants and in coupling signals collected from different imaging modalities, preventing the development of ‘digital twin’ models²¹ for plants.

Moreover, monitoring wood degradation using multimodal imaging necessitates proper registration and the identification of specific signatures for structural and physiological traits

of interest. Such signatures are not well described on wood but are mandatory to deploy automatic quantitative approaches. Experts should perform a preliminary conjoint analysis of multimodal imaging signals together with manual pixel-wise annotation of wood degradation before deploying automatic morphological phenotyping methods based on tissue segmentation and machine learning.

To address these issues, we developed an end-to-end workflow for in-vivo phenotyping of internal woody tissues condition, based on multimodal and non-destructive 3D imaging, and assisted by AI-based automatic segmentation. This approach was applied to vine imaging datasets acquired in a clinical imaging facility. 3D images were acquired in five different modalities (X-rays for structure, three MRI parameters for function, and serial sections for expertise) on entire plants, and combined by an automatic 3D registration pipeline²². Based on serial section annotations, structural and physiological signatures characterizing early- and late-stages of wood degradation were identified in each imaging modality. Then a machine-learning model was trained to automatically classify tissue condition based on the multimodal imaging data, achieving high performance in distinguishing healthy and sick tissues. We could, therefore, perform an accurate quantification of intact, degraded, and white rot compartments within entire vine trunks. We also evaluated the contribution and efficiency of each imaging technique for tissue detection. Finally, we studied the relationships between external foliar

Automatic tissues segmentation. (a) AI-based image segmentation using multimodal signals. (b) Comparison of all possible imaging modality combinations for their effectiveness (F1-scores) in tissue detection. (c) Effectiveness of tissue detection at lower imaging resolutions (using four modalities). (d) 3D reconstructions highlighting the extent and localization of the degraded and white rot compartments in four vines.

The same validation protocol was used to compare the effectiveness of all possible combinations of imaging modalities for tissue detection (Fig. 3b): the most efficient combination was [T1-w, T2-w, X- ray] for detection of intact (F1 = 93.9 ± 3.4%) and degraded tissues (90.5 ± 3.2%); and [T1-w, X-ray] for white rot (93.0 ± 5.1%). Interestingly, the X-ray modality alone reached almost similar scores for white rot detection. In general, slightly better results (± 0.5%) were obtained without considering the PD-w modality, likely due to its lower initial resolution.

The classifier was finally applied to the whole dataset. Under the assumption of the proper generalization of our model, we utilized all available data, including computations on non-ground truth pixels during inference, to compute robust statistics and compare tissue contents in different vines. Considering the entire classified dataset (46.2 million voxels), mean signal values significantly declined between intact and degraded tissues (− 19.3% for X-ray absorbance; and − 57.3%, − 86.3% and − 71.3% for MRI T1-w, T2-w, and PD-w, respectively); and between degraded and white rot (− 56.0% for X-ray absorbance; and − 36.8%, − 76.8% and − 64.2% for MRI T1-w, T2-w, and PD-w, respectively) (Fig. 2e and Table S4).

With the increasing deployment of X-ray and NMR devices on phenotyping tasks, in-field imaging has become accessible, but at a heavy cost in terms of image quality and resolution. We challenged our method by training and evaluating the classifier at coarser resolutions, ranging from 0.7 to 10 mm per voxel (Fig. 3c). Results proved our approach maintained correct performances even at 10 mm (F1 ≥ 80% for intact; ≥ 70% for degraded and white rot), while a human operator can no longer recognize any anatomical structure or tissue class at this resolution.

These results confirmed the wide range of potential applications and the complementarity of the four imaging modalities. Combining medical imaging techniques and an AI-based classifier, it was possible to segment intact, degraded, and white rot compartments automatically and nondestructively inside the wood. This represents an important breakthrough in their visualization, volume quantification and localization in the entire 3D volume of the vines (Fig. 3d).

Deciphering the relationship between inner tissue composition and external symptomatic histories: A step toward a reliable in-situ diagnosis

Non-destructive detection of GTDs in vineyards is currently only possible through observing foliar symptoms and vine mortality. Numerous studies are based on these proxies for phenotyping. Foliage is usually screened at specific periods of the year when Esca or Eutypa dieback symptoms occur, and this screening is usually repeated for one or several years. However, new leaves are produced each year, and symptoms may not recur in following years, making any diagnosis hazardous at best, and attempts to correlate external and internal impacts of GTDs unsuccessful. Here, foliar symptoms were recorded yearly for twenty years since the plot was planted in 1999 (Fig. 4a). Together with the accurate quantification and localization of degraded and non-degraded compartments in trunks, this allowed more advanced investigations.