Rapid, reliable, robust classification of freshwater mussel haemocytes using deep learning into two major classes

Abstract

Bivalve mussels are crucial indicator organisms in toxicological, ecological, and immunological studies due to their long life cycle, tolerance capabilities, fast recovery, rapid growth, filter-feeding behaviour, and importance in fisheries and aquaculture. They respond to rapid changes in their environment. Haemocytes, which provide innate immunity and composition changes, are affected by ecological conditions, disease, and environmental stressors, leading to immune system and histopathology changes. Haemocyte numbers and types vary considerably among bivalve groups, with granulocytes and hyalinocytes being the main cell types. There are significant knowledge gaps, controversies, and missing data regarding haemocyte morphology and classification in freshwater mussels and marine bioindicator organisms. The development of unified classification and nomenclature systems to describe bivalve haemocytes has become imperative, particularly considering the threats posed to public health and ecosystems by climate change and emerging diseases. Existing studies on this subject are predominantly based on the morphological properties of cell images. The proposed study employs a convolution neural network (CNN) as a deep learning algorithm to classify mussel haemocytes, yielding objective and robust results. It is demonstrated that the two cell types can be characterized with a high degree of accuracy. Among the CNN models evaluated, the highest accuracy value 0.9666 ± 0.0372, the best recall 0.9666 ± 0.0567, the best precision 0.9696 ± 0.0503, the best F1 score 0.9665 ± 0.0375 and the best Area Under Curve (AUC) 0.9666 ± 0.0372 were obtained with VGG-16. The findings demonstrate that CNN possesses the capacity to objectively identify mussel haemocytes, and hence make a contribution to the field.