Detection of Microconidia in Microscopy Images of Fusarium oxysporum f. sp. cubense Using Image Processing Techniques and Neural Networks

• Main Authors: Abigan, Erinn Giannice T, Cajucom, Luis Gabriel A, Ong, Josh Daniel L, Abu, Patricia Angela R, Estuar, Ma. Regina Justina E
• Format: text
• Published: Archīum Ateneo 2020
• Subjects: image segmentation; microscopy; neural networks; soil; fluorescence; feature extraction; diseases; Fusarium oxysporum f. sp. cubense; microscopy image analysis; convolutional neural networks; microconidia detection; Computer Sciences; Databases and Information Systems; Plant Sciences
• Online Access: https://archium.ateneo.edu/discs-faculty-pubs/224; https://ieeexplore.ieee.org/document/9334941
• Institution: Ateneo De Manila University

Fusarium oxysporum f. sp. cubense (Foc) is a soilborne fungus and the causative agent of the deadly Fusarium wilt disease in banana plants. Left alone, the fungus is able to survive for years and infect multiple plants through the soil. External symptoms only manifest in late stages of infection, with the destruction of all plants within a 7.5 meter radius of the diseased through burning being the only way to eradicate the fungus. Foc Tropical Race 4 (TR4) is capable of infecting the widely used Cavendish cultivars, threatening global banana production. It is imperative then that Foc be detected as soon as possible. To achieve this, the study endeavors to detect microconidia, a reproductive structure of the Foc species, in microscopy images of stained soil specimen under three microscopy configurations using image processing techniques and convolutional neural networks (CNNs). The networks were built using the ResNet-50 architecture, and results were validated via Gradient-weighted class activation mapping (Grad-CAM). The network classifying fluorescent images achieved the highest accuracy with 95.24%, followed by bright field images with 94.94 %, all (bright field, dark field, and fluorescent) images with 93.75 %, and lastly, dark field images with 92.86%. Grad-CAM results indicate the networks are able to identify Foc structures and correctly distinguish clean from Foc-infected images. This study contributes towards the early detection of Foc, and is a step toward mitigating the threat it presents.