Deep Learning-Based Classification of In Vivo Embryonic Development Stages in <i>Ceratitis capitata</i> (Wiedemann) (Diptera: Tephritidae) Using Real-Time Microscopy Images

Tuncyurek T., AKÇURA M., Genc H.

JOURNAL OF EXPERIMENTAL ZOOLOGY PART B-MOLECULAR AND DEVELOPMENTAL EVOLUTION, 2026 (SCI-Expanded, Scopus)

Özet

Accurate staging of embryonic development is crucial for applied studies that require precise determination of embryonic age, such as developmental biology research and phenotype-based assessments using microinjection. However, the determination of embryonic development stages in the Mediterranean fruit fly, Ceratitis capitata, still largely relies on manual morphological assessment, which is labor-intensive and susceptible to observer bias. In this study, we evaluated deep learning models for the automatic classification of embryonic development stages in C. capitata using time-lapse microscopy images. Embryos were monitored in vivo at 25 degrees C +/- 1 degrees C, and high-resolution 4 K phase-contrast images (3840 & times; 2160 pixels) were acquired at hourly intervals throughout the developmental process (50 h). After quality control, a balanced dataset of 3000 images representing six development stages was created, including blastoderm formation (BF), early gastrulation (EG), germband elongation (GE), germband retraction (GR), dorsal closure (DC), and muscle movement (MM). Four pre-trained architectures (EfficientNetV2M, EfficientNetV2S, ResNet50, and DenseNet121) were standardized and compared under the same conditions. Model performance was evaluated using overall accuracy, sensitivity, specificity, precision, F1 score, ROC, precision-recall analysis, confusion matrices, and t-SNE visualization. With the single training-validation-test split used here, EfficientNetV2M reached a test accuracy of 86.22% and a macro-F1 of 86.12%, and ResNet50 gave comparable results. As each model was trained only once, these small differences are treated descriptively rather than as evidence of statistical superiority. Classification was most reliable for the BF, EG, and MM stages, whereas more confusion arose at the intermediate stages, where morphological features overlap. Overall, these findings show that deep learning offers a practical framework for the time-dependent classification of embryonic development stages in C. capitata and could support future work in embryology and applied entomology.