High throughput saliency-based quantification of grape powdery mildew at the microscopic level for disease resistance breeding.

Abstract

Imaging-based high throughput phenotyping (HTP) systems have demonstrated promising solutions to enhance genetic understanding of grapevine powdery mildew (PM) resistance and have accelerated PM-resistant cultivar breeding. The accuracy and throughput of extracting phenotypic traits from images are still the bottleneck of modern HTP systems, especially at the microscopic level. The goal of this study was to develop a saliency-based processing pipeline for the quantification of PM infection in microscopic images and comprehensively evaluate its performance for genetic analyses. An input image was segregated into subimages that were classified as infected or healthy by a pretrained CNN classifier. Saliency maps from the classification were generated post-hoc and used for the quantification of PM infection in the input image at the pixel level without the use of mask annotations. A total of seven phenotypic traits were extracted from images collected for a biparental population. Experimental results showed that optimal combinations of convolutional neural network and saliency methods achieved strong measurement correlations (r = 0.74 to 0.75) with human assessments at the image patch level, and the traits calculated by the saliency-based processing pipeline were highly correlated (r = 0.87 to 0.88) with reference PM infection ratings at the leaf image level. The high quantification accuracy of the saliency-based pipeline led to the increased explanation of phenotypic variance and reliable identification of quantitative trait loci. Therefore, the saliency-based processing pipeline can be used as an effective and efficient analysis tool for PM disease research and breeding programs in the future, especially agricultural and life science studies requiring microscopic image analysis.