Exploring Artificial Intelligence for age estimation in cephalopod beaks

Age data are crucial for understanding the life cycle and the population dynamics of the species for assessment and sustainable exploitation of cephalopods. The formation of daily growth increments has been validated in several hard structures (e.g. statoliths, beaks, shells), and is currently used in many species. The microstructure analysis is time‐consuming since it implies having to identify several hundred narrow increments per individual. We explored how Deep Learning can assist in detecting the growth increments in beaks. An analysis via Convolutional Neural Networks was performed on juvenile and adult beaks to achieve the best AI model. A set of beak images of Octopus vulgaris (juvenile and adult stages) was used to train neural networks (NNs) using the image segmentation tool Segment.ai ©Nikon Nis.ai software. A high‐resolution camera provided quality images to draw the growth increments following different tracking patterns. We tested the following segmentation parameters for each NN: tracking pattern of the growth increments, number of training images, number of iterations and re‐training. We assessed the accuracy of the NN:a) Quantitatively based on the training loss or error percentageb) Qualitatively by running each NN on a set of unsegmented images to visually check and compare the results. Main conclusions: Tracking the light part of the increment and 1500 training iterations (accuracy-training time) showed the best results. Training loss is not always in line with the visual checking