Investigation of cardinality classification for bacterial colony counting using explainable artificial intelligence

Automatic bacterial colony counting is a highly sought-after technology in amodern biological laboratory since it can eliminate manual counting efforts. Previouswork in automatic colony counting has observed that MicrobiaNet, whichis the best-performing cardinality classification model for colony counting to thebest of our knowledge, has difficulty distinguishing colonies of three or more individuals. However, it is unclear if this is due to properties of the data togetherwith inherent characteristics of the MicrobiaNet model. Using explainable artificialintelligence, we provide a detailed analysis of MicrobiaNet’s performance.Our results demonstrate that the high visual similarity across different classesin the data set is the key issue hindering further enhancement of the model’sperformance. Future work should focus on developing models that incorporatethis factor or on exploring density estimation methods for colony counting. Ouranalysis has potential wider impacts on applications of neural network classifiers.