Scale-Dependent Performance of YOLOv12 andYOLOv11 for Automated Blood Cell Detection: A Controlled Benchmark on BCCD

Problem Statement

Automated detection of red blood cells (RBCs), white blood cells (WBCs), and platelets in peripheral blood smears is critical for improving the speed and reproducibility of hematological analysis. While deep learning, particularly the YOLO (You Only Look Once) family, has become dominant for this task, the performance benefits of the newly released YOLOv12 over its predecessor, YOLOv11, remain unclear. Specifically, YOLOv12 introduces an "attention-centric" design featuring Area Attention (A2) modules and Residual Efficient Layer Aggregation Network (R-ELAN) blocks, theoretically offering better context reasoning for morphologically distinct and overlapping cells. However, no study had previously conducted a controlled, scale-by-scale benchmark comparing all five variants (nano to extra-large) of YOLOv12 against YOLOv11 on the standard Blood Cell Count and Detection (BCCD) dataset under identical experimental conditions.

Methodology

The authors conducted a rigorous, controlled benchmark to eliminate confounding variables often found in comparative studies.

• Dataset: The study utilized the BCCD v3 dataset, comprising 364 microscopic blood smear images. A fixed, reproducible 70/15/15 train/validation/test split was applied (random seed 42), resulting in 255 training images, 55 validation images, and 54 test images.
• Data Augmentation: To address the small training set size, offline augmentation was applied exclusively to the training split. Seven transforms (flips, rotations, crops, HSV jitter) were used to select two augmentations per image, expanding the training set from 255 to 765 images. Validation and test sets remained un-augmented.
• Models: All ten models were evaluated: five YOLOv12 variants (n, s, m, l, x) and five YOLOv11 variants (n, s, m, l, x).
  • YOLOv11: Utilizes C3k2 backbone blocks and a C2PSA (Convolutional block with Parallel Spatial Attention) neck.
  • YOLOv12: Replaces the C2PSA neck with an Area Attention (A2) module (applying self-attention within localized regions) and uses R-ELAN blocks in the backbone.
• Training Protocol: All models were trained for 100 epochs on a single NVIDIA A100-SXM4-40 GB GPU using Ultralytics 8.x and PyTorch 2.10. Hyperparameters (optimizer, momentum, weight decay) were identical across all runs, with learning rates adjusted only by scale. Models were initialized with COCO-pretrained weights.
• Evaluation Metrics: Performance was assessed on the held-out test set using mAP@0.5, mAP@0.5:0.95, per-class Average Precision (AP), inference latency, and GFLOPs.
• Statistical Analysis: To determine if performance differences were due to sampling noise or genuine architectural advantages, the authors applied McNemar's exact test to paired per-image detection outcomes (detected vs. missed) for each scale variant.

Key Results

1. Overall Performance and Efficiency

• Mean Performance: Across all five scales, the mean mAP@0.5 was nearly identical between the two families: 91.40% ± 0.91% for YOLOv12 and 91.38% ± 0.69% for YOLOv11.
• Best Single Model: YOLOv12n achieved the best trade-off for edge deployment, with the highest mAP@0.5 (92.51%), the lowest latency (1.8 ms), and the lowest compute cost (6.5 GFLOPs).
• Per-Class Performance: WBC detection was saturated across all models (AP > 98%). RBC detection showed a slight edge for YOLOv11 (mean AP 85.77% vs. 85.35%), while Platelet detection was highest at the extra-large scale for both (92.95%).

2. Scale-Dependent Statistical Differences
While aggregate mAP was similar, McNemar's exact test revealed significant, non-monotonic differences in recall based on model scale:

• Nano Scale: YOLOv12n significantly outperformed YOLOv11n in recall (94.53% vs. 91.38%, p = 0.0008).
• Small Scale: No statistically significant difference was found (p = 0.1034).
• Medium and Large Scales: The trend reversed. YOLOv11 significantly outperformed YOLOv12 in recall at the medium scale (89.60% vs. 82.08%, p < 0.0001) and large scale (90.29% vs. 85.23%, p < 0.0001).
• Extra-Large Scale: Both architectures converged to identical recall (96.44%, p = 1.0000) and identical mAP@0.5 (92.14%).

3. Error Analysis

• Consensus Failures: An analysis of objects missed by at least 8 of the 10 models identified 21 "consensus failures." These were predominantly RBCs partially occluded by other cells or platelets with low contrast. This suggests that remaining errors are driven by image quality and annotation ambiguity (e.g., overlapping cells) rather than specific architectural limitations of either YOLOv11 or YOLOv12.
• Training Dynamics: Both architectures demonstrated smooth loss reduction and stable training across all scales, with no evidence of divergence.

Key Contributions

1. First Controlled Scale-by-Scale Benchmark: This is the first study to evaluate all five scale variants of YOLOv12 against YOLOv11 on the BCCD dataset under strictly identical conditions (data split, augmentation, hyperparameters, and hardware).
2. Revealing Scale-Dependent Architecture Efficacy: The study demonstrates that YOLOv12 does not uniformly outperform YOLOv11. Instead, the advantage of the attention-centric design is highly scale-dependent, favoring YOLOv12 only at the nano scale and YOLOv11 at medium and large scales.
3. Statistical Rigor: By applying McNemar's exact test to recall, the authors uncovered significant performance gaps that were masked by aggregate mAP metrics, providing a more nuanced view of detection sensitivity.

Significance and Claims

The paper claims that model selection for automated blood cell detection should not rely solely on aggregate mAP scores but must consider the specific scale of the model and the operational requirements (latency vs. recall).

• For Edge Deployment: YOLOv12n is identified as the superior choice, offering the highest mAP, lowest latency, and a statistically significant recall advantage over its YOLOv11 counterpart.
• For Recall-Critical Applications: At medium and large scales, YOLOv11 is preferable, as it provides statistically significant recall improvements (up to 7.5 percentage points) over YOLOv12.
• Diminishing Returns: At the extra-large scale, both architectures converge, suggesting that simply increasing model capacity will not resolve the remaining detection errors caused by cell overlap and low contrast. Future gains likely depend on improved annotation consistency and handling of overlapping erythrocytes rather than architectural changes alone.

The authors conclude that the introduction of Area Attention in YOLOv12 provides a measurable benefit only at the smallest model capacity, while the convolutional design of YOLOv11 remains more robust for intermediate scales in this specific domain.