Coherent lensless diffraction imaging via multicore fiber for label-efficient microplastics characteristic discrimination

The widespread exposure of microplastics (MPs) to environmental and biological systems makes the rapid identification of their damage status a critical task. This deformation influences how MPs migrate, degrade, and interact with organisms, thereby shaping their environmental fate and ecological risk. However, traditional bright-field microscopy suffers from degraded performance under defocused and low-contrast conditions, and benchtop systems are also difficult to deploy in situ in confined spaces. To address these challenges, we propose a lightweight method combining multicore fiber bundle-based coherent defocused imaging with self-supervised representation learning. When only an extremely small number of annotations are provided, semi-supervised constraints, including weighted updates and soft penalties are introduced to enhance cluster-class alignment. This scheme eliminates the need for holographic reconstruction and features a thin, flexible probe, making it suitable for in-situ and online detection in confined environments. The results demonstrate that combining physically enhanced coherent imaging with self-supervised features enables robust differentiation between damaged and undamaged MPs at low annotation cost, and provides a feasible path for on-site deployment.