Accurate segmentation of Natural Killer (NK) cells is critical for the quantitative analysis of peritoneal immune cells. In this study, we compared several convolutional neural network (CNN) architectures—FCN32s, FCN16s, FCN8s, FCN-ResNet50, U-Net, and SegNet—using 6,560 time-lapse images derived from endometriosis and non-endometriosis datasets. Model performance was evaluated through ten-fold cross-validation based on morphological features (area, perimeter, circularity) and quantitative metrics, including Intersection over Union (IoU), precision, recall, and average precision (AP). FCN32s and FCN16s exhibited substantial spatial information loss, resulting in oversimplified masks. FCN8s and FCN-ResNet50 achieved improved segmentation but showed limitations in recall and stability. By contrast, U-Net and, most notably, SegNet outperformed all FCN-based models, demonstrating higher accuracy, lower variability, and more robust segmentation. SegNet achieved the best performance across all evaluation criteria. These results suggest that while FCN-based models provide a useful baseline, U-Net and SegNet offer superior generalization, structural adaptability, and reliability for NK cell morphological analysis. The integration of AI-based segmentation with conventional manual analysis presents a practical strategy to enhance efficiency and reproducibility in immunological studies.