Automated biometric analysis of early fish life stages via instance segmentation using Mask-RCNN

Biometric measurements of early life stages of fish are a critical component of studies of fish development. This is a time consuming and laborious manual task. For large data sets such manual labour rapidly becomes limiting factor in the experimental space that can be covered and accuracy of the human-generated output.

We present a method that utilities transfer learning of Mask R-CNN, by Facebook AI Research, for reliable and accurate instance segmentation of images, enabling automated geometrical measurements of features and objects within microscope images. The key advantage of this approach is that objects of interest can be identified and quantified even if the image has a complex background or the lighting conditions or focus are varying, making the automated biometric measurements robust. Using classical image processing on the segmented images from the neural network we can obtain accurate biometric measurements such as larvae length, body area, yolk sac area, myotome height, eye diameter eye-to forehead distance and heart cavity area.

The successful implementation of this technique for studies of early fish larvae development has equated to a reduction in analysis cost of up to 98% and expanded the capability to analyse vastly larger datasets than was previously possible. We present a demonstration of this method in quantifying biometry in fish larvae exposed to produced water and crude oil in the laboratory, in order to understand how spills of these toxicants in the marine environment may impact local populations of spawning products of Atlantic cod. Based on our image analyses we concluded that exposure to produced water caused significant effects on fish larvae, e.g. smaller larvae size, larger yolk sacs and smaller eyes, in addition to craniofacial deformations. The availability of this tool in providing accurate, consistent and high-volume information on the internal structure of small organisms opens many new possibilities for expanding the scope of developmental and effect studies within marine biology and ecotoxicology.