Accurate segmentation of individual thyroid cells is a prerequisite for cell feature analysis and reliable cancer staging classification. However, Diff-Quick stained cytology images present significant challenges: frequent misclassification of malignant cells and erythrocytes and substantial cell overlap hindering boundary detection. To address these issues, we propose a novel two-stage pipeline. This approach enhanced the efficiency-optimized nnU-Net v2 for rapid foreground-background separation, enabling efficient instance segmentation of overlapping cells and reducing erythrocyte misclassification. The results evaluated on a dataset with multiple thyroid cancer stages show that our method reduced erythrocyte false positives and improved accuracy over the best post-processed baseline while cutting inference time. These findings demonstrate the practical utility of our pipeline for automated Diff-Quick thyroid cytology image segmentation within real-world clinical workflows.