M-Count: an application that uses machine learning object detection and color thresholding to count settled mussel larvae

Abstract

Quantification of biofouling is a complex task that often involves counting both isolated and tightly packed groups of organisms on a test surface. Mussel larvae, which settle as both individuals and as clumps, are an important fouling organism because adult mussel colonies form via the settlement of larvae, making the settlement or repellence of mussel larvae a good indicator of a surface's antifouling performance. Manual quantification methods are time-consuming, and existing automatic machine learning-based methods are poorly suited for use by coding non-experts and often lack the ability to detect both isolated and grouped organisms in one workflow. The objective of this work was to develop a machine learning-based application that is user-friendly and well-suited to the quantification of biofouling. We developed M-Count, an application that combines a neural network object detection model for individual organism detection and a color thresholding algorithm for grouped organism detection. M-Count was demonstrated on the quantification of mussel larvae on sample surface images obtained from a previously developed mussel larvae fouling assay. This study revealed three important characteristics of M-Count: speed, consistency, and accuracy. The primary benefit of M-Count is its speed, being 60× faster than manual counting. The secondary benefit of M-Count is consistency, as it performs the same task repeatedly without bias. Finally, these benefits are obtained while maintaining good accuracy, with the normalized average maximum residual being 0.220 for M-Count and 0.209 for manual counting.