Pixel-Based Machine Learning and Image Reconstitution for Dot-ELISA Pathogen Diagnosis in Biological Samples.

Area of Science:

• Biotechnology
• Machine Learning in Diagnostics
• Plant Pathology

Background:

• Serological methods are crucial for diagnosing pathogen infections across species.
• Dot-ELISA (DE) is a cost-effective and sensitive diagnostic tool, widely used in epidemiology.
• Human interpretation of DE results can be subjective due to overlapping colorations, limiting diagnostic accuracy.

Purpose of the Study:

• To develop an objective and accurate method for evaluating Dot-ELISA (DE) results using machine learning.
• To overcome the limitations of human visual interpretation in DE assays.
• To enable precise classification of positive and negative samples for accurate pathogen diagnosis.

Main Methods:

• A supervised machine learning approach was employed, utilizing a multivariate logistic regression model.
• The model was trained on RGB pixel data from scanned DE outputs of known infected and uninfected samples.
• The algorithm predicts infection status based on pixel probabilities in scanned DE images.

Main Results:

• The developed machine learning model accurately classifies DE outputs, providing unambiguous diagnostic results.
• The method allows for adjustable cutoffs to balance false positive and false negative rates.
• Successful application demonstrated for diagnosing *Lettuce big-vein associated virus*.

Conclusions:

• Machine learning-based image analysis offers a versatile and impartial solution for DE assay interpretation.
• This approach enhances diagnostic accuracy and reliability in various epidemiological applications.
• The method translates unique pathogen antigens into a universal color-based diagnostic language.