Identifying Candida Species Using Machine Learning
Disciplines
Bacterial Infections and Mycoses | Biomedical Informatics | Biotechnology | Cell Anatomy | Engineering | Life Sciences | Microbial Physiology | Other Immunology and Infectious Disease | Pathogenic Microbiology
Abstract (300 words maximum)
The genus Candida is a vast collection of more than 200 species of yeast. Some of these are commensals that live harmoniously with the human body, while others are pathogenic and can cause various diseases. Discovered in 2009, C. auris is a pathogenic species of Candida that can rapidly spread in clinical settings and cause severe disease. C. auris is associated with increased morbidity compared to other Candida species and has a mortality rate of 30-60% in hospitalized patients when paired with other illnesses. Typically, Candida spp. are difficult to differentiate, and many diagnostic methods are prone to misidentification. Current identification methods involve germ tube, enzymic, and commercial automated tests or emerging molecular typing techniques. These methods are either time-consuming, require advanced lab equipment, or have a high misidentification rate for C. auris. This project focuses on developing a time and cost-effective diagnostic platform using accessible laboratory equipment. Our approach involves taking microscope images of cell cultures of Candida spp. and applying machine-learning algorithms to find morphological differences in species.
We discuss two deep learning models that output a classification label given an image. The first, Artificial Neural Network (ANN) takes in continuous cell metrics processed from images, including cell area and cell circularity. The second, Convolution Neural Network (CNN) takes in image data and applies various filters to identify features of an image. Our previous work on binary-classification showed a precision of 80% between C. auris and C. haemulonii. With our ANN model on five species, we achieved 70% precision with several species. We will discuss these two deep learning models applied to five species of Candida, in seek of greater than 95% precision of all classes. This methodology allows for the adaptive identification of many Candida species and has the potential to be expanded to other micro-identification domains.
Academic department under which the project should be listed
CSM - Molecular and Cellular Biology
Primary Investigator (PI) Name
Christopher Cornelison
Identifying Candida Species Using Machine Learning
The genus Candida is a vast collection of more than 200 species of yeast. Some of these are commensals that live harmoniously with the human body, while others are pathogenic and can cause various diseases. Discovered in 2009, C. auris is a pathogenic species of Candida that can rapidly spread in clinical settings and cause severe disease. C. auris is associated with increased morbidity compared to other Candida species and has a mortality rate of 30-60% in hospitalized patients when paired with other illnesses. Typically, Candida spp. are difficult to differentiate, and many diagnostic methods are prone to misidentification. Current identification methods involve germ tube, enzymic, and commercial automated tests or emerging molecular typing techniques. These methods are either time-consuming, require advanced lab equipment, or have a high misidentification rate for C. auris. This project focuses on developing a time and cost-effective diagnostic platform using accessible laboratory equipment. Our approach involves taking microscope images of cell cultures of Candida spp. and applying machine-learning algorithms to find morphological differences in species.
We discuss two deep learning models that output a classification label given an image. The first, Artificial Neural Network (ANN) takes in continuous cell metrics processed from images, including cell area and cell circularity. The second, Convolution Neural Network (CNN) takes in image data and applies various filters to identify features of an image. Our previous work on binary-classification showed a precision of 80% between C. auris and C. haemulonii. With our ANN model on five species, we achieved 70% precision with several species. We will discuss these two deep learning models applied to five species of Candida, in seek of greater than 95% precision of all classes. This methodology allows for the adaptive identification of many Candida species and has the potential to be expanded to other micro-identification domains.