Automated detection of `Ca. Liberibacter asiaticus' infection in citrus using immune tissue prints and machine learning

Authors: Shao, Jonathan; DING, FANG - Huazhong Agricultural University; FU, SHIMIN - Southwest University; Hartung, John

Submitted to: Book Chapter
Publication Type: Book / Chapter
Publication Acceptance Date: 3/20/2020
Publication Date: 5/11/2021
Citation: Shao, J.Y., Ding, F., Fu, S., Hartung, J.S. 2021. Automated detection of `Ca. Liberibacter asiaticus' infection in citrus using immune tissue prints and machine learning. Plant Diseases and Food Security in the 21st Century, Plant Pathology in the 21st Century, Springer Nature pp. 215-230. https://doi.org/10.1007/978-3-030-57899-2_10.
DOI: https://doi.org/10.1007/978-3-030-57899-2_10

Interpretive Summary: Huanglongbing (HLB) or citrus greening disease has caused unprecedented losses to the citrus industry in Florida since its discovery there in 2005. Sweet orange and grapefruit production has been reduced by 80 and 96% respectively. HLB is also present in Texas and California and is a global problem for citriculture. Sustained and productive research has greatly increased the understanding of HLB but control measures remain insufficient. HLB is caused by a bacterial infection that is spread by an insect. Diagnosis of HLB is by identification of the pathogen in infected trees. Infected trees are often destroyed to prevent further spread of the disease. Identification of the pathogen is difficult and we have developed a new antibody based method for this purpose that ends with a digital photo of plant tissue. An expert can then determine if the plant was infected with HLB or not. A problem with this technique is that an expert is required to interpret the image, the result is a clear yes or no, and humans have biases and become tired. To address this, we have used advanced artificial intelligence methods to create computer programs that can interpret these digital photos. This provides unbiased results, eliminates a tired or bored expert, and provides a confidence score which reflects the certainty of the result. Our results will be used by professional diagnosticians and others tasked with the identification of this pathogen in citrus trees in areas where HLB is a problem.

Technical Abstract: Huanglongbing (HLB) or citrus greening disease has caused unprecedented losses to the citrus industry in Florida since its’ discovery there in 2005. HLB is caused by an infection by the unculturable member of the alpha-proteobacteria ‘Ca. Liberibacter asiaticus’ (Clas). Diagnosis of HLB is by identification of the pathogen in infected trees and infected trees are often destroyed to prevent further spread of the disease. Identification of the pathogen is generally done by PCR-based methods that are expensive, labor intensive and requires isolation of DNA from tree samples prior to assay. Even worse, the nonuniform distribution of the pathogen in citrus trees means that there will be a high rate of ‘false negative’ test results if the tissue selected for assay from an infected tree did not contain the pathogen. We have previously developed a direct tissue blot immunoassay (DTBIA) assay to supplement PCR-based testing. The DTBIA scales well to large numbers of samples and produces digital images of cross sections of plant tissue that have been serologically probed to reveal the presence of Clas. The assay works well but relies on a human expert to classify the image as infected or not infected. The process can be tedious, and subject to inconsistency and bias. To address this, we trained a convolutional neural network (CNN) based on Inception V3 architecture available in Tensorflow to interpret the images. Sets of curated images were prepared from healthy or diseased citrus petioles or stems tested by DTBIA and used to train the CNN models. After the models were developed they the curated sets of positive and negative training images were combined and submitted as unknowns to the CNN models for interpretation. The model based on images of petiole sections gave slightly better results than the stem model, with 81% of images classified in agreement with the curator at more than 90% confidence. Similar results were obtained when images that were not used to train the models were submitted as unknowns. There were two instances noted in which the model incorrectly classified an image of a very plainly infected DTBIA as uninfected with 84% and 74% confidence. Errors such as this may be difficult to eliminate for technical reasons and some curator oversight will be necessary. The models will be useful in the classification of DTBIA results for the protection of the citrus industry.