<![CDATA[Manual cell counting from haemocytometer images is slow, subjective, and operator-dependent, especially in unstained brightfield microscopy where cell boundaries and viability-related morphology are difficult to distinguish. Although prior cell detection models have mainly been evaluated on stained or fluorescence images, systematic comparisons between fine-tuned detectors and zero-shot cell segmentation models remain limited for unstained brightfield haemocytometer images. This study presents a controlled 2×2 factorial benchmark of patchification and augmentation across five detection approaches, with variance-decomposition analysis and comparison of fine-tuned versus zero-shot deployment modes. Using 24 unstained brightfield RAW 264.7 macrophage images with 6,307 polygon-level annotations, including 28.8% dead cells, we evaluated four preprocessing scenarios under six-fold stratified cross-validation. Faster R-CNN, Mask R-CNN, and YOLOv11n-Seg were fine-tuned within each fold, whereas Cellpose and StarDist were applied zero-shot. Grid-aligned patchification improved bounding-box mAP50 by 2.6–8.4× across all fine-tuned architectures (paired Wilcoxon p = 0.016, Cohen’s d > 3). A 2×2 ANOVA attributed 99.2–99.4% of explained variance to patchification, while augmentation and interaction effects each contributed less than 0.1%, suggesting that performance gains were driven mainly by scale rescaling rather than sample count. On patchified data, fine-tuned models converged to 85.5–86.4% mAP50. YOLOv11n-Seg achieved the highest mAP50-95 of 51.1%, with 6× faster inference and 17× fewer parameters. In contrast, zero-shot Cellpose and StarDist reached only 45.3–51.2% class-agnostic F1@0.5. These findings show that structure-aware patchification is critical for reliable cell detection in this modality.]]>
