Phenotypic identification of locust embryonic development using deep learning and optical coherence tomography

Locusts are valuable model organisms whose studies of embryonic development elucidate genetic mechanisms of insect development and provide a theoretical foundation for pest management and evolutionary biology. Optical coherence tomography (OCT), with high-resolution, noninvasive, real-time imaging, is an effective tool for monitoring locust embryogenesis and RNA interference phenotypes. However, OCT image analysis remains difficult due to subtle daily variations, time-consuming manual annotation, and inherent subjectivity. To overcome these challenges, we developed a deep learning framework for embryonic phenotype analysis and staging. The framework integrates ResNet34 with transfer learning, AdamW optimization, and class activation mapping (Grad-CAM) visualization, highlighting morphological structures relevant for discrimination across days and constructing a reference atlas of locust embryonic development. Experiments show the framework achieves 98.8% accuracy across 13 days, greatly reducing manual workload and improving staging consistency. This study enables automated phenotype recognition, offering theoretical and practical value for insect embryology and pest management.