In-vivo flow cytometry is a noninvasive real-time diagnostic technique that facilitates continuous monitoring of cells without perturbing their natural biological environment, which renders it a valuable tool for both scientific research and clinical applications. However, the conventional approach for improving classification accuracy often involves labeling cells with fluorescence, which can lead to potential phototoxicity. This study proposes a label-free in-vivo flow cytometry technique, called dynamic YOLOv4 (D-YOLOv4), which improves classification accuracy by integrating absorption intensity fluctuation modulation (AIFM) into YOLOv4 to demodulate the temporal features of moving red blood cells (RBCs) and platelets. Using zebrafish as an experimental model, the D-YOLOv4 method achieved average precisions (APs) of 0.90 for RBCs and 0.64 for thrombocytes (similar to platelets in mammals), resulting in an overall AP of 0.77. These scores notably surpass those attained by alternative network models, thereby demonstrating that the combination of physical models with neural networks provides an innovative approach toward developing label-free in-vivo flow cytometry, which holds promise for diverse in-vivo cell classification applications.