Automatic Blood Cell Detection (ABCD)

• The paper demonstrates that integrating CBAM, ASFF, and CIoU into YOLOX significantly enhances detection precision and localization in blood smear images.
• It employs a deep-learning object detection framework to identify RBCs, WBCs, and platelets, addressing challenges like overlap and scale variations.
• Experimental results on the BCCD dataset show improved mAP and FPS compared to baseline methods, indicating robust performance in practical clinical scenarios.

Searching arXiv for recent and directly relevant blood-cell detection papers to ground the article. Automatic Blood Cell Detection Method (ABCD) denotes a deep-learning approach for detecting blood cells in microscopic blood smear images, introduced as an improved YOLOX-based detector for the three principal blood cell categories: red blood cells (RBCs), white blood cells (WBCs), and platelets (Hasen et al., 25 Jul 2025). In the cited formulation, ABCD is motivated by the clinical importance of blood-cell examination for disorders such as anemia, malaria, leukemia, and other blood disorders, and by the limitations of manual smear review, which is described as time-consuming, inefficient, subjective, and error-prone (Hasen et al., 25 Jul 2025). Within the broader literature on automated hematology image analysis, ABCD is best understood as an object-detection system rather than a segmentation-free classifier alone: it takes a blood smear image, applies preprocessing and augmentation, processes the image through an improved YOLOX architecture, and outputs class labels and bounding boxes for RBCs, WBCs, and platelets (Hasen et al., 25 Jul 2025). This places it in a lineage of microscopy-oriented detectors ranging from Faster R-CNN-based WBC localization (Xia et al., 2019) to YOLO-family blood-cell detectors (Liu et al., 2023, Wu et al., 4 Jun 2025), while remaining distinct from crop-level blood-cell classifiers that assume isolated cells as input (Gavas et al., 2021, Mondal et al., 2024).

1. Definition and task scope

ABCD is formulated as an automatic blood cell detection method for microscopic blood smear images, with the explicit target categories RBC, WBC, and platelets (Hasen et al., 25 Jul 2025). Its stated objective is to reduce the workload and variability associated with manual blood smear inspection by replacing direct visual review with a deep-learning detector that can identify multiple cell types in a single image (Hasen et al., 25 Jul 2025).

The method is framed around the practical difficulties of blood smear imagery. RBCs are described as numerous and densely packed, WBCs as sparse and irregular, and platelets as very small and easy to miss (Hasen et al., 25 Jul 2025). The paper further states that traditional image-processing approaches struggle in regions where cells overlap, cluster, or vary substantially in shape and size, which motivates the use of a deep-learning detector rather than a purely handcrafted segmentation-and-feature pipeline (Hasen et al., 25 Jul 2025).

A useful contrast within the literature is that some blood-cell studies address only a subset of the full ABCD problem. For example, Faster R-CNN-based work on microfluidic point-of-care imaging focuses on WBC detection only, with binary labeling of WBC vs other objects (Xia et al., 2019). Other studies benchmark deep CNNs for single-cell classification after cropping, rather than full-scene detection (Gavas et al., 2021). ABCD differs in that it is explicitly a multi-class detector over whole microscopic images, not merely a recognition model for already isolated cells (Hasen et al., 25 Jul 2025).

2. Architectural formulation

The baseline architecture of ABCD is YOLOX, chosen because it is described as fast and accurate, with an anchor-free design and a decoupled head (Hasen et al., 25 Jul 2025). The baseline detector comprises three parts: backbone, neck, and head. The backbone is CSPDarknet/CSPNet53, the neck uses PAFPN, and the head is decoupled so that classification and bounding-box regression are handled by separate branches (Hasen et al., 25 Jul 2025).

ABCD modifies YOLOX in three places. First, it inserts the Convolutional Block Attention Module (CBAM) into the backbone. Second, it introduces Adaptively Spatial Feature Fusion (ASFF) into the neck. Third, it replaces the baseline IoU loss with CIoU loss for bounding-box regression (Hasen et al., 25 Jul 2025). The resulting pipeline is described as follows: start from a blood smear image, apply preprocessing and augmentation, feed the image to the improved YOLOX, extract features in the backbone enhanced by attention, fuse multi-scale features in the neck with adaptive weighting, and output final class labels and bounding boxes for RBCs, WBCs, and platelets through the YOLO detection heads (Hasen et al., 25 Jul 2025).

The role of CBAM is to refine backbone features so that the network focuses on diagnostically relevant cell structures while suppressing background and clutter (Hasen et al., 25 Jul 2025). If the input feature map is

$F \in \mathbb{R}^{C \times H \times W},$

then the sequential attention refinement is

$$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$

where $M_c(F)\in \mathbb{R}^{C\times 1\times 1}$ is channel attention and $M_s(F')\in \mathbb{R}^{1\times H\times W}$ is spatial attention (Hasen et al., 25 Jul 2025). Channel attention is defined by

$$M_c(F)=\sigma\big(\text{MLP}(\text{avgpool}(F))+\text{MLP}(\text{maxpool}(F))\big),$$

and spatial attention by

$$M_s(F)=\sigma\big(f^{7\times 7}([\text{avgpool}(F);\text{maxpool}(F)])\big)$$

(Hasen et al., 25 Jul 2025). The paper argues that this is especially useful in blood-cell imagery because RBCs may be densely distributed and overlapping, WBCs may appear with irregular shapes, and platelets occupy very small image regions (Hasen et al., 25 Jul 2025).

The role of ASFF is to improve multi-scale fusion for objects of different apparent size. With three pyramid feature maps $X_1$, $X_2$, and $X_3$, the paper gives the example of fusion at level 3:

$$ASFF_3 = X_{1\rightarrow 3}\otimes \alpha_3 + X_{2\rightarrow 3}\otimes \beta_3 + X_{3\rightarrow 3}\otimes \gamma_3,$$

where resized features are combined with learned adaptive spatial weights (Hasen et al., 25 Jul 2025). The text does not provide an explicit normalization equation for $$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$0, $$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$1, and $$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$2, so only the existence of spatially adaptive weighting can be stated directly. This suggests that ABCD treats platelet-scale and RBC-scale evidence differently at different pyramid levels, rather than relying on fixed fusion.

The third change is the use of CIoU loss for box regression. The paper gives:

$$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$3

with

$$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$4

The printed expression for $$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$5 is partially corrupted, but the accompanying text states that $$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$6 measures aspect-ratio consistency between the predicted and ground-truth boxes (Hasen et al., 25 Jul 2025). The paper explains the three terms directly: overlap reward, center-distance penalty, and aspect-ratio consistency enforcement (Hasen et al., 25 Jul 2025). In crowded blood-cell scenes, this is intended to improve localization and accelerate convergence.

3. Data basis and experimental protocol

ABCD is evaluated on the BCCD blood cell dataset, described as an open-source set of annotated blood smear images created and gathered by MIT and made publicly available on Roboflow (Hasen et al., 25 Jul 2025). According to the paper, the dataset contains 364 images with 4,888 annotations, and each image has size $$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$7 (Hasen et al., 25 Jul 2025). The annotated classes are RBC, WBC, and platelets (Hasen et al., 25 Jul 2025).

The original dataset is split into training, validation, and test sets in a 7:2:1 ratio (Hasen et al., 25 Jul 2025). Because annotation is labor-intensive and the dataset is relatively small, the paper applies several augmentation techniques: flipping, hue adjustment, rotation, and exposure modification (Hasen et al., 25 Jul 2025). After augmentation, the reported numbers of images used for training, validation, and testing are 1800, 175, and 87, respectively (Hasen et al., 25 Jul 2025). The paper also mentions preprocessing such as resize, labeling, and contrast adjustment, though exact preprocessing parameters are not elaborated in detail (Hasen et al., 25 Jul 2025).

Implementation was performed in Python 3.8 using PyTorch under Windows, with CUDA 11.4.125 (Hasen et al., 25 Jul 2025). The reported hardware configuration is an Intel Core i7-10750H CPU, 16 GB memory, and an NVIDIA GeForce GTX 1650 Ti GPU (Hasen et al., 25 Jul 2025). All compared models were trained for 50 epochs with an initial learning rate of 0.01 and a 5-epoch warmup; the optimizer was Adam with default parameter settings (Hasen et al., 25 Jul 2025). The architecture figure shows the model input as $$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$8 (Hasen et al., 25 Jul 2025).

Several details are explicitly absent from the paper. It does not specify a batch size, weight decay, momentum, or a full learning-rate schedule beyond the initial learning rate and warmup (Hasen et al., 25 Jul 2025). This limits exact reproduction. A related observation appears elsewhere in the literature: blood-cell detection papers often provide strong detector-level results but incomplete implementation detail, including sparse reporting of thresholds and software settings (Xia et al., 2019, Liu et al., 2023).

4. Evaluation and empirical performance

ABCD reports Precision, Recall, F1-score, mAP@0.5, [email protected]–0.9, parameter count in millions, and FPS as its primary metrics (Hasen et al., 25 Jul 2025). The definitions given in the paper are

$$F' = M_c(F) \otimes F,\qquad F'' = M_s(F') \otimes F',$$9

$M_c(F)\in \mathbb{R}^{C\times 1\times 1}$0

$M_c(F)\in \mathbb{R}^{C\times 1\times 1}$1

and

$M_c(F)\in \mathbb{R}^{C\times 1\times 1}$2

(Hasen et al., 25 Jul 2025).

On the BCCD test data, ABCD achieves 90.80% precision, 92.28% recall, 91.65% F1-score, 95.49% [email protected], 86.89% [email protected]–0.9, 9.45M parameters, and 73.6 FPS (Hasen et al., 25 Jul 2025). Compared to baseline YOLOX, which achieves 92.69% [email protected], 63.48% [email protected]–0.9, 8.94M parameters, and 69.8 FPS, ABCD improves [email protected] by 2.8 percentage points, improves [email protected]–0.9 by 23.41 percentage points, and increases speed by 2.9% (Hasen et al., 25 Jul 2025).

The class-wise performance reported by the paper is as follows:

Class	Precision	Recall	AP50
RBC	98.90%	97.60%	92.60%
WBC	99.90%	99.40%	99.94%
Platelets	99.30%	93.20%	96.49%

These values indicate that the model is strongest on WBCs, then platelets, and somewhat lower on RBCs, which the authors attribute to the dense distribution and overlapping characteristics of RBCs (Hasen et al., 25 Jul 2025). This interpretation is consistent with the visual structure of smear images and with other studies emphasizing overlap and small-object difficulty as the dominant detection bottlenecks (Liu et al., 2023, Wu et al., 4 Jun 2025).

The paper compares ABCD with several alternative detectors on BCCD:

Method	[email protected]	[email protected]–0.9	FPS
Faster R-CNN	78.94%	51.94%	11.2
YOLOv3	90.39%	56.94%	39.4
YOLOv4	91.36%	58.30%	32.7
YOLOv5	92.30%	61.82%	67.2
YOLOX	92.69%	63.48%	69.8
ABCD	95.49%	86.89%	73.6

Among these compared detectors, ABCD is reported as both the most accurate and the fastest (Hasen et al., 25 Jul 2025). The paper also states that ABCD surpasses prior blood-cell-specific methods such as TE-YOLOF (91.90% mAP), FED (89.86% mAP), YOLOv3-SPP (88.6% mAP), AYOLOv5 (93.3% mAP), and ISE-YOLO (85.7% mAP, 34.5 FPS) (Hasen et al., 25 Jul 2025).

5. Ablation evidence and interpretation

The ablation study is central to understanding what ABCD contributes beyond baseline YOLOX. Starting from baseline YOLOX, the reported progression is:

Variant	[email protected]	[email protected]–0.9	FPS
YOLOX baseline	92.69%	63.48%	69.8
+ Augmentation	93.84%	84.28%	71.2
+ CBAM	94.62%	86.04%	70.6
+ ASFF	94.86%	86.58%	72.2
+ CIoU	95.49%	86.89%	73.6

The largest single gain in the strict localization metric comes from augmentation, with [email protected]–0.9 increasing by 20.01 points (Hasen et al., 25 Jul 2025). The paper interprets this as evidence for the importance of larger and more varied training data (Hasen et al., 25 Jul 2025). This suggests that, in the BCCD regime, data scarcity and limited variability are at least as important as architecture.

CBAM then provides a meaningful additional improvement, especially for smaller objects (Hasen et al., 25 Jul 2025). ASFF further improves results by refining multi-scale fusion (Hasen et al., 25 Jul 2025). CIoU adds a smaller but consistent gain and is said to improve regression quality and convergence (Hasen et al., 25 Jul 2025). The cumulative pattern implies that ABCD is not one isolated module but a composed detector in which preprocessing/augmentation, attention-guided representation learning, scale-aware fusion, and improved box regression all contribute.

There is a numerical inconsistency noted in the paper between some textual ablation summaries and the final tables. One paragraph mentions 95.18% [email protected] and 87.2% [email protected]–0.9, whereas Tables 3 and 6 report 95.49% and 86.89% (Hasen et al., 25 Jul 2025). The final tables are the clearest quantitative source. This suggests that the strongest factual basis for ABCD’s headline performance is the final tabulated result rather than intermediate prose.

6. Position within the broader literature

ABCD sits within a rapidly expanding body of blood-cell detection methods based on deep object detectors. A compact historical reference point is Faster R-CNN-based WBC detection in microscopy for microfluidic point-of-care devices, which used transfer learning on the BCCD Dataset, a small dataset consisting of 364 images, with 50 images for training and 314 images for testing, and reported 98.4% overall accuracy and 1.3% miss rate for WBC detection (Xia et al., 2019). That system, however, addressed WBC detection only and did not perform multi-class detection over RBCs and platelets (Xia et al., 2019).

Later YOLO-family methods increasingly emphasized lightweight deployment and small-object sensitivity. ADA-YOLO, built on YOLOv8 with an Adaptive Head, reports on BCCD that it improves YOLOv8 in mAP@50 and mAP@50-95 while using 8.7 MB and 9.4 GFLOPs, compared with 26.9 MB and 35.1 GFLOPs for YOLOv8 (Liu et al., 2023). MS-YOLO, built from YOLOv11-n, reports 97.4% mAP@50 on the CBC benchmark, with explicit focus on overlapping cells and very small platelets (Wu et al., 4 Jun 2025). These developments indicate that blood-cell detection has become a specialized subfield of medical object detection in which architectural changes are driven by three recurring image properties: overlap, scale disparity, and small-object recall.

ABCD shares these concerns directly. Its use of CBAM parallels the broader use of attention mechanisms in medical detectors (Liu et al., 2023). Its ASFF-based scale fusion aligns with other multi-scale strategies used for blood-cell imagery (Wu et al., 4 Jun 2025, Chen et al., 2024). At the same time, its task scope differs from leukocyte-only detectors such as MFDS-DETR, which focuses on leukocyte detection and reports AP = 79.7 on a private five-class leukocyte dataset but is not intended as a full RBC/WBC/platelet detector (Chen et al., 2024).

A common misconception in this area is to treat high single-cell classification performance as equivalent to full automatic blood cell detection. This is not correct. Studies such as "Deep CNNs for Peripheral Blood Cell Classification" classify cropped single-cell images into eight categories and report 99.51% ensemble accuracy, but they do not solve localization in full smear scenes (Gavas et al., 2021). ABCD, by contrast, is a whole-image detector. Another misconception is to equate all blood-cell “detection” papers with object detection; some papers use the word for crop-level malaria classification or segmentation-first pipelines (Mondal et al., 2024). ABCD specifically belongs to the bounding-box detection tradition (Hasen et al., 25 Jul 2025).

7. Limitations, interpretation, and significance

The paper is explicit about several limitations. First, the dataset is relatively small, which may limit generalization (Hasen et al., 25 Jul 2025). Second, the additions of CBAM and ASFF increase model complexity and computational requirements relative to baseline YOLOX, even though the final model remains efficient (Hasen et al., 25 Jul 2025). Third, RBC detection remains more challenging than WBC and platelet detection because RBCs are densely packed and often overlap (Hasen et al., 25 Jul 2025).

A broader limitation, also visible in related work, is that reported performance is usually confined to a small number of public datasets under consistent imaging conditions (Xia et al., 2019, Liu et al., 2023, Chen et al., 2024). ABCD is reported only on BCCD (Hasen et al., 25 Jul 2025). There is no external multi-center validation, no analysis across different staining protocols or microscope manufacturers, and no calibration or uncertainty study for clinical deployment. This suggests that the detector should be understood as a strong benchmarked method rather than a clinically validated hematology instrument.

The paper nonetheless has clear practical significance. By retaining the fast, anchor-free YOLOX framework and strengthening it through CBAM in the backbone, ASFF in the neck, and CIoU for box regression, ABCD offers a concrete recipe for real-time blood-cell detection under conditions of small objects, overlap, and heterogeneous scales (Hasen et al., 25 Jul 2025). The reported 73.6 FPS and 95.49% [email protected] indicate that the method is aimed not merely at offline analysis but at deployment contexts where throughput matters (Hasen et al., 25 Jul 2025).

A plausible implication is that ABCD’s strongest contribution lies not in introducing an entirely new detector family, but in showing that a carefully targeted combination of attention-guided feature extraction, adaptive multi-scale fusion, and improved localization loss can produce large gains in strict localization quality on blood smear imagery. The especially large increase in [email protected]–0.9 over baseline YOLOX suggests that the method improves not only whether cells are found, but how precisely they are localized (Hasen et al., 25 Jul 2025). In blood-smear analysis, where crowded scenes and tiny platelets make localization error clinically consequential, that distinction is significant.