Analysis of the Effects of Microscopy Techniques on the Performances of Convolutional Neural Network Architectures in Microscopic Fusarium Microconidia Detection

Fusarium wilt, a disease afflicting banana plants, is caused by the soilborne fungus Fusarium oxysporum f. sp. cubense or Foc. Upon entering a plant, Foc attacks the vascular system of the host. Infected plants show symptoms late, when Foc may already be widespread both in the plant and in the plantation, and often do not recover. The recommended way to deal with Foc, once detected, is to quarantine and burn plants and soil within a 7.5 meter radius, resulting in the loss of a few crops at best and whole plantations at worst. One way to prevent Foc from wreaking havoc is to detect it early on. The goal of the study is to develop and compare convolutional neural networks (CNNs) identifying microconidia, a fungal structure, in microscopy images with three microscopy techniques. Four CNN architectures classified images into either Clean or Foc (microconidia present). In terms of accuracy, by microscopy technique, CNNs classifying bright field (BF) images consistently yielded the highest, followed by those classifying fluorescent (FL) images, then All images, and lastly, dark field (DF) images. By architecture, the ResNet-50 CNNs consistently performed the best, followed by ResNet-101, then VGG-19 with batch normalization, and AlexNet. In terms of prediction time, by microscopy technique, the All images networks took 3–4 times longer than the BF, DF, and FL networks. By architecture, AlexNet consistently took the least time, followed by VGG-19 with batch normalization, then ResNet-50, and ResNet-101.