Automatic localization and segmentation of adherent microplastics in optical micrographs based on improved YOLOv5 and adaptive perceptual UNET 3+++

Abstract

Rapid and accurate measurements of microplastic residues in biological tissues are crucial for determining exposure levels and studying microplastic invasion mechanisms. However, manual extraction of microplastics from micrographs is labor-intensive, especially for small, irregularly shaped adhesive particles from the whole blood of carp with complex backgrounds. This study proposed a detection and segmentation approach of a two-level fusion network with a variable universe, which can simultaneously achieve accurate segmentation and counting of adherent microplastics of different shapes and scales. In the target detection model, YOLOv5 serves as a prototype, and the RS adaptive structure is introduced, which improves the accuracy by 4.9 %, the mAP by 4.4 %, and the computation amount is reduced by 22.8 % compared with the traditional YOLOv5 model. In the segmentation model, with UNet3 + as the prototype, ECA (Efficient Channel Attention) is introduced in the last two layers of the full-scale inter-skip connection, and DANet(Dual Attention Network) is integrated into the last layer of the encoder, the average F1 score achieved on a dataset of 2831 microplastic images was 99.65 %. This result effectively meets the requirement for accurate extraction of each microplastic particle. Overall, the extraction of adherent microplastics from the whole blood of carp validates the potential of optical microimaging and two-level fusion network-based detection and segmentation methods for the extraction of adherent microplastics from human blood.