Automated Detection and Cropping of Hyphae in Microscopic Images of Various Fungi

Fusarium oxysporum f. sp. cubense is a soil-borne fungi that has become a major threat to the current banana industry. The presence of this fungi can destroy entire plantations if not detected and stopped early enough. The purpose of this study is to create a Convolutional Neural Network (CNN) that can detect hyphae in microscopic images. By detecting hyphae, the presence of fungi in the soil can be confirmed. To create a model that can detect hyphae, a dataset of various microscopic images of fungi was sorted into hyphae images and non-hyphae images (labeled as others). Four subsequent datasets were created from this, namely: (1) bright field, (2) dark field, (3) fluorescent, and (4) all microscopy techniques. Pretrained ResNet34 and ResNet152 models were used for each of the datasets and the use of heatmaps on these models was done to analyze how the models looked for hyphae. The ResNet34 model achieved accuracies of 86.38% for bright field, 87.31% for dark field, 88.37% for fluorescent, and 87.60% for all microscopy techniques. The ResNet152 model achieved accuracies of 87.97% for bright field, 86.79% for dark field, 89.37% for fluorescent, and 87.69% for all microscopy techniques. Additionally, to improve the accuracy even further, automated cropping using edge detection and contour detection was done on the datasets to create cropped photos of hyphae. This resulted in average test accuracies of 87.17% for bright field, 86.90% for dark field, 91.22% for fluorescent, and 89.99% for all microscopy techniques. Generally, fluorescent consistently performed the best, but the heatmaps generated from each model show that hyphae can also be detected using the other microscopy techniques. This study can act as a steppingstone for future studies involving the classification of fungi through hyphae and other features.