A Comparison of Three Automated Root-Knot Nematode Egg Counting Approaches Using Machine Learning, Image Analysis, and a Hybrid Model

Abstract

Meloidogyne spp. (root-knot nematodes [RKNs]) are a major threat to a wide range of agricultural crops worldwide. Breeding crops for RKN resistance is an effective management strategy, yet assaying large numbers of breeding lines requires laborious bioassays that are time-consuming and require experienced researchers. In these bioassays, quantifying nematode eggs through manual counting is considered the current standard for quantifying establishing resistance in plant genotypes. Counting RKN eggs is highly laborious, and even experienced researchers are subject to fatigue or misclassification, leading to potential errors in phenotyping. Here, we present three automated egg counting models that rely on machine learning and image analysis to quantify RKN eggs extracted from tobacco and sweet potato plants. The first method relied on convolutional neural networks trained using annotated images to identify eggs (M. enterolobii R2 = 0.899, M. incognita R2 = 0.927, M. javanica R2 = 0.886), whereas a second contour-based approach used image analysis to identify eggs from their morphological characteristics and did not rely on neural networks (M. enterolobii R2 = 0.977, M. incognita R2 = 0.990, M. javanica R2 = 0.924). A third hybrid model combined these approaches and was able to detect and count eggs nearly as well as human raters (M. enterolobii R2 = 0.985, M. incognita R2 = 0.992, M. javanica R2 = 0.983). These automated counting protocols have the potential to provide significant time and resource savings annually for breeders and nematologists and may be broadly applicable to other nematode species.