Detection and identification of insect processed animal proteins in microscopic images using deep learning.

The introduction of terrestrial invertebrate processed animal proteins (PAPs) from insects in animal feed has introduced challenges for official controls in the European Union. Currently, their detection relies only on light microscopy and is entirely dependent on the expertise of trained operators. Recent studies have highlighted two primary problems regarding insect PAP detection: specificity issues due to the misidentification of non-insect structures as insect derived ones, and the difficulty in categorizing a wide morphological diversity of particles resulting from the grinding of whole insect larvae. To address these limitations, a proof of concept was developed involving automated microscopic image analysis and classification by deep learning. A classification pipeline integrating a YOLO object detection model with an improved ConvNeXt architecture coupled with transformer blocks was tested. After supervised learning, training and validation of the model, obtained results allowed to successfully distinguish insect particles from other non-insect materials, and to discriminate PAPs particles derived from Tenebrio molitor and Hermetia illucens achieving average Matthews Correlation Coefficient of 0.89 and accuracies exceeding 95% for both tasks. The robust performance and the rapid processing speed of the model enabled to couple it to a digital microscope for real-time automated particle classification, including confidence scores for individual particle predictions. This artificial intelligence model for microscopic identification of complex structures opens new perspectives for official controls achieving performance scores never reached before. It provides to microscopists an optimisation of image analysis and a valuable decision-support tool in complex feed matrices where human expertise is frequently challenged.