A hierarchical multi-scale neural network framework for schistocyte detection in blood smears

Research gaps methods

This research work is designed to directly address the pervasive limitations in current blood cell detection literature, which include: (1) coarse-grained dataset annotations that overlook critical pathological cells (e.g., schistocytes); (2) inadequate handling of micro-scale targets (2–50 μm) and morphological complexity; and (3) suboptimal trade-offs between real-time performance and accuracy. As illustrated in Fig. 1, our workflow begins with a collaboration with medical institutions to collect high-quality blood smear images, followed by the construction of the first comprehensive dataset (SBSD) containing six cell categories—erythrocytes, platelets, granulocytes, lymphocytes, monocytes, and schistocytes—through manual annotation and pathological expert verification. This dataset directly fills the annotation gap by providing fine-grained labels for pathological cells, enabling robust model training.

The dataset construction process leverages the backbone network of MCS-Net, enhanced by a single-head self-attention (SHA) mechanism (Yun & Ro, 2024) to strengthen local feature capture, addressing the lack of detail in existing models for irregular cell morphologies. Image data processed through the feature pyramid network is then fed into our novel Multi-Scale Conv Attention (MSCA) module, which models local details and global contextual dependencies through multi-branch dilated convolutions and attention fusion. This achieves effective feature fusion for multi-scale targets, directly mitigating the literature’s deficiency in handling size disparities (e.g., platelets at 2–3 μm vs. leukocytes at 15–20 μm). Finally, through optimization of MCS-Net’s neck network and the Cross-Scale Dynamic (CSD) detection head, coupled with the introduction of normalized Wasserstein distance (NWD) loss, we significantly improve micro-target detection performance and localization accuracy. This integrated approach ensures a balance between accuracy and computational efficiency, overcoming the real-time limitations of prior works.

Overall architecture

As illustrated in Fig. 1, our MCS-Net workflow consists of four main stages:

• (1)

  Data Acquisition and Annotation: We collaborated with medical institutions to collect high-quality blood smear images. The SBSD dataset was constructed through manual annotation by pathologists, ensuring fine-grained labels for six cell types. This stage addresses annotation gaps by including pathological cells like schistocytes.

• (2)

  Backbone Network with Single-Head Self-Attention (SHA): The backbone network is enhanced with an SHA mechanism to strengthen local feature capture for irregular morphologies. This step improves detail extraction, often lacking in existing models.

• (3)

  Multi-Scale Feature Processing with MSCA Module: Feature maps from the backbone are processed through a feature pyramid network (FPN) and fed into our MSCA module. This module uses parallel dilated convolutional branches (e.g., dilation rates 1, 2, 3) to capture multi-scale features, followed by attention fusion to combine local and global contexts. This design mitigates size disparity issues by adaptively handling targets from 2 to 50 μm.

• (4)

  Detection and Optimization with CSD Head and NWD Loss: The neck network is optimized with a CSD detection head that dynamically fuses features across scales. The NWD loss function is introduced to enhance localization accuracy for micro-targets (e.g., platelets and schistocytes), reducing errors in dense clusters. This ensures high accuracy without sacrificing real-time performance.

Figure 1 visually summarizes this workflow, highlighting the integration of SHA, MSCA, CSD, and NWD components. This architecture is specifically designed to overcome the limitations of prior works by providing a unified solution for multi-scale, pathological cell detection.

Dataset morphological characteristics

Figure 5 delineates the diameter distribution of six cell types in the SBSD dataset using kernel density estimation (KDE) curves, providing critical insights into morphological characteristics essential for algorithm design and clinical validation.

RBCs: Peak at 8.0 μm (μ = 8.0 ± 1.0 μm), aligning with physiological biconcave disc morphology. The narrow standard deviation (σ = 1.0 μm) reflects uniformity, justifying RBCs as a stable reference for size calibration; PLT: Sharp peak at 2.5 μm (μ = 2.5 ± 0.5 μm), consistent with their anucleate fragment nature. The minimal spread (σ = 0.5 μm) highlights their distinct size separation from other categories; granulocytes (GRAN): Largest mean diameter (μ = 17.1 ± 2.8 μm), attributable to lobulated nuclei and cytoplasmic granularity. The broad KDE curve (spanning 12–22 μm) underscores size heterogeneity across subtypes (neutrophils, eosinophils, basophils); Schistocytes: Overlap with PLTs and monocytes (MONOs) at 4.0 μm (μ = 4.0 ± 1.5 μm). Their KDE curve (green) exhibits a bimodal distribution (peaks at 3.5 and 5.5 μm), suggesting two subpopulations: small fragments (mechanical shear-induced) and larger irregular remnants (intrinsic defects).

Schistocyte-PLT Ambiguity: Both schistocytes (4.0 ± 1.5 μm) and PLTs (2.5 ± 0.5 μm) occupy the 2–6 μm range, with 24% overlap in KDE densities. This explains baseline model misclassifications and validates the need for NWD-based localization to distinguish subtle shape differences; Lymphocyte-Monocyte Differentiation: Lymphocytes (LYM: μ = 12.6 ± 2.5 μm) and monocytes (MONO: μ = 17.1 ± 2.8 μm) show partial overlap (14–16 μm) (Sabattini et al., 2010), likely due to transitional forms in pathological states (e.g., activated lymphocytes); Multi-Scale Detection Necessity: The dataset spans 2.5 μm (PLTs) to 17.1 μm (GRANs), demanding hierarchical feature fusion (MSCA module) to capture both micron-level fragments and large cells; Schistocyte/MONO size variability (σ = 1.5 μm) NWD to mitigate IoU’s sensitivity to minor positional shifts; Schistocyte Quantification: The 4.0 ± 1.5 μm distribution matches ICSH diagnostic criteria for thrombotic microangiopathy (≥1% schistocytes with 2–5 μm size), confirming SBSD’s utility in DIC/MAHA screening (Iba et al., 2019).

Cellular abundance in SBSD dataset

Figure 6A: Main plot (logarithmic scale): The bar chart illustrates cell counts spanning four orders of magnitude (10²–10⁵), dominated by erythrocytes (RBCs: 77.4k, 93.5%). Pathological markers, highlighted in blue, reveal schistocytes (2.0k, 2.4%), exceeding the diagnostic threshold (≥1%) for thrombotic microangiopathy per ICSH guidelines (Zini & De Cristofaro, 2019). The granulocyte-lymphocyte ratio (GLR = 3.7, GRAN: 3.1k vs. LYM: 2.0k) aligns with bacterial infection profiles (normal GLR: 1.5–3.0) (Naess et al., 2017), validating the dataset’s inclusion of infection-related samples. Figure 6B: Enlarged linear-scale plot: Resolves low-abundance populations, including granulocytes (3.1k ± 15%) and monocytes (200) (Huang, Cai & Su, 2019), critical for detecting rare pathological events (e.g., monocyte activation in sepsis). Error bars on granulocytes reflect biological variability across infection subtypes.

Through its morphological diversity annotations and cellular abundance distribution design, the SBSD dataset not only reflects the physiological-pathological continuum of blood components but also provides multi-scale, high-complexity detection scenarios for algorithmic development. Its construction directly addresses clinical demands, such as determining diagnostic thresholds for schistocytes in disseminated intravascular coagulation (DIC). It establishes a validation foundation for innovative modules (MSCA, CSD, NWD) in the MCS-Net model, thereby advancing blood cell detection from coarse-grained classification to pathological subtype recognition.

Multi-scale conv attention global branch

The global self-attention branch serves as the core component for capturing global dependency relationships within the input features. Its architecture integrates the classical self-attention mechanism, where the input feature map is first passed through depthwise separable convolutional layers to generate three sets of vectors: Q, K, and V. Global correlations are computed via normalized dot product operations, and a learnable temperature parameter is introduced to modulate the sharpness of the attention distribution, as formulated in Eq. (1):

(1) $$AttentionScore\left(Q,K,V\right)=Softmax\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}\cdot \tau }\right).$$

Here, $d$ denotes the dimensionality of the key vectors, which scales the dot product values to prevent gradient explosion, and $\tau$ is a learnable parameter that dynamically optimizes the attention weight distribution.

The value vectors $V$ are aggregated through a weighted summation based on the attention weights, yielding a globally contextualized feature representation as formalized in Eq. (2):

(2) $$F_{out1}=Con{v}_{1\times 1}\left(AttentionScore\left(Q,K,V\right)\cdot V\right)+X^{\mathrm{\prime }}.$$

Here, $\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}$ denotes a 1 × 1 convolutional layer, and $X^{\mathrm{\prime }}$ represents the dimensionally reduced input features. This weighted summation enables the model to prioritize spatially correlated regions relevant to the current cellular context (e.g., schistocytes) and adjacent pathological zones (e.g., platelet aggregation areas).

Multi-scale conv attention local branch

The primary objective of the local feature extraction branch is to capture fine-grained spatial details within input features through multi-scale convolutional operations. In designing this branch, we integrate strategies for multi-scale feature extraction and cross-channel interaction enhancement. Specifically, parallel convolutional kernels with varying receptive fields—such as standard 3 × 3 kernels, kernels with a dilation rate of 2, and kernels with a dilation rate of 3—are applied to mimic the multi-focal switching behavior employed by pathologists when examining cellular morphology. At each stage of feature processing, cross-channel information interaction is further enhanced through channel dimension adjustment and reordering.

The input to the local branch is first processed through a 1 × 1 convolutional layer, ${\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}=Con{\mathrm{v}}_{1\times 1}\left(\mathrm{R}\mathrm{S}\left({\mathrm{X}}^{\mathrm{\prime }}\right)\right)$, followed by RS (channel dimension adjustment and rearrangement operations).

Subsequently, the local branch performs multi-scale feature extraction and fusion via parallel convolutional kernels, including standard convolution, dilated convolution with a rate of 2, and dilated convolution with a rate of 3. The convolutional formulations for each branch are specified in Eqs. (3) and (4):

(3) $$\{\begin{array}{c}{F}_1=Con{v}_{3\times 3}^{padding=1}\left(F_{conv}\right)\hfill \\ F_2=Con{v}_{3\times 3}^{padding=2,dilation=2}\left(F_{conv}\right)\hfill \\ F_3=Con{v}_{3\times 3}^{padding=3,dilation=3}\left(F_{conv}\right)\hfill \end{array}$$

(4) $$F_{out1}=Con{v}_{1\times 1}\left(AttentionScore\left(Q,K,V\right)\cdot V\right)+X^{\mathrm{\prime }}).$$

The final local output is further augmented via a residual connection to ${\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}$, as formalized in Eq. (5):

(5) $$F_{out2}=F_{fused}+F_{conv}.$$

Three-branch fusion

As illustrated in the architectural diagram, the MSCA employs residual connections to preserve low-level details from the original input features. This mechanism, when integrated with the local and global branches, achieves complementary information integration, effectively mitigating information degradation during deep network training. Specifically, the local branch enhances edge-texture features (e.g., serrated cell membrane edges), while the global branch models long-range dependencies (e.g., spatial co-occurrence relationships between platelets and schistocytes). Additionally, the original residual structure retains chromatic information from staining patterns. The final output is formulated as Eq. (6):

(6) $$F_{out2}=F_{fused}+F_{conv}.$$

Here, ${\mathrm{F}}_{\mathrm{o}\mathrm{u}\mathrm{t}1}$ denotes the global features output by the global attention branch, ${\mathrm{F}}_{\mathrm{o}\mathrm{u}\mathrm{t}2}$ represents the local features generated by the local convolutional branch, and $\mathrm{X}$ corresponds to the input feature map (original features).

Cross-scale interaction

A list of multi-scale feature maps ${\mathrm{x}}_{\mathrm{l}\mathrm{i}\mathrm{s}\mathrm{t}}={\left\{{\mathrm{F}}_{\mathrm{i}}\right\}}_{\mathrm{i}=1}^{\mathrm{N}}$ is input, where ${\mathrm{F}}_{\mathrm{i}}\in {\mathrm{R}}^{\mathrm{B}\times {\mathrm{C}}_{\mathrm{i}}\times {\mathrm{H}}_{\mathrm{i}}\times {\mathrm{W}}_{\mathrm{i}}}$ denotes the feature map at the i scale. Subsequently, channel-wise descriptors are generated via Global Average Pooling (GAP) (Lin, Chen & Yan, 2013), as formulated in Eq. (7):

(7) $$S_i=GAP\left(F_i\right)={\displaystyle \frac{1}{H_i{W}_i}\sum _{h=1}^{H_i}\sum _{w=1}^{W_i}{F}_i^{\left(c\right)}\left(h,w\right).}$$

This operation compresses each scale-specific feature into a channel-wise descriptor $S_i\in R^{B\times C_i}$, preserving contextual information across scales.

The descriptors from all scales are concatenated and fed into a fully connected (FC) network, as formalized in Eq. (8):

(8) $$concat\mathrm{\_}feat=Concat\left(S_1,S_2,\dots S_N\right).$$

The concatenated descriptors are fed into the FC network, which generates dynamic parameters through dimensionality reduction and nonlinear transformation, as formalized in Eq. (9):

(9) $$W,b=FC\left(concat\mathrm{\_}feat\right)=W_2\cdot ReLU\left(W_1\cdot concat\mathrm{\_}feat+b_1\right)+b_2$$where the dynamic parameter vector $\mathrm{W}\in {\mathrm{R}}^{\mathrm{B}\times 2\mathrm{N}}$ is partitioned into two components: $\mathrm{\alpha }\in {\mathrm{R}}^{\mathrm{B}\times \mathrm{N}}$ (weight parameters) and $\mathrm{\beta }\in {\mathrm{R}}^{\mathrm{B}\times \mathrm{N}}$ (bias parameters).

The formulation for generating weights and bias parameters is formalized in Eq. (10):

(10) $$\alpha =Softmax(W[:,:N]),\beta =Tanh(W[:,:N])$$where the Softmax function ensures ${\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{\alpha }}_{\mathrm{i}}=1$ to normalize the weight distribution, while the Tanh activation constrains the bias parameters b within [−1, 1] to prevent feature shift caused by numerical instability.

Finally, weighted summation and bias adjustment are applied to the features at each scale, as formalized in Eq. (11):

(11) $$F_i^{\mathrm{\prime }}={\alpha }_i\otimes F_i+{\beta }_i$$where $\otimes$ denotes element-wise multiplication. This operation enhances critical scales (e.g., the P3 layer for schistocyte detection) through channel-wise dynamic scaling, while the bias term preserves the original staining characteristics.

DynamicHead

The DynamicHead module introduces DynamicConv, which innovatively decouples the expansion of the parameter space (i.e., increasing the number of expert kernels, K) from computational costs. By integrating a multi-expert mechanism with dynamic routing principles, it achieves an optimal balance between enhanced model capacity and inference speed (Han et al., 2023). As illustrated in Fig. 10 to address lightweight requirements, the classification branch combines Depthwise Separable Convolution with DynamicConv (Howard et al., 2017), further minimizing computational overhead.

The dynamic convolution kernel is generated through a linear combination of multiple expert convolution kernels, as formalized in Eq. (12):

(12) $$W_{dynamic}=\sum _{k=1}^K{\alpha }_k\cdot W_k,{\alpha }_k=sigmoid(W_{routing}\cdot GAP(x)).$$

Here, a linear transformation is applied to map the input channel dimension to the number of experts, and the sigmoid activation function ensures non-sparse weight allocation.

Gaussian distribution modeling for bounding boxes

In the SBSD dataset, most cells exhibit non-rectangular morphologies (e.g., platelets with biconvex discoid or elliptical shapes), which inherently misalign with axis-aligned rectangular bounding boxes. Directly applying rectangular bounding boxes in SBSD inevitably encapsulates partial staining background regions, thereby introducing localization inaccuracies. To address this, we model bounding boxes as two-dimensional Gaussian distributions, where the central region (corresponding to the actual cellular structure) is assigned higher pixel weights, while weights decay towards the edges. The mathematical formulation for the inscribed ellipse is provided in Eq. (14):

(14) $${\displaystyle \frac{{\left(x-c_x\right)}^2}{{\sigma }_x^2}+{\displaystyle \frac{{\left(y-c_y\right)}^2}{{\sigma }_y^2}=1.}}$$

Here, (cx, cy) denotes the center coordinates of the rectangle, while ${\mathrm{\sigma }}_{\mathrm{x}}$ and ${\mathrm{\sigma }}_{\mathrm{y}}$ represent the semi-axis lengths along the x-axis and y-axis of the rectangle, respectively.

For any axis-aligned bounding box defined by its center coordinates $\left(c_x,c_y\right)$, width $w$, and height $h$, we parameterize it as a 2D Gaussian distribution $N\left(\mu ,\mathrm{\Sigma }\right)$ through the following steps:

(1) Mean Vector $\mathit{\mu }$: The mean is set to the center of the bounding box, placing the peak of the Gaussian distribution at the box’s center: $\mu =\left[\begin{array}{c}{c}_x\\ c_y\end{array}\right]$.

(2) Covariance Matrix Σ: The covariance matrix is set to a diagonal matrix, where the diagonal elements are derived from the width and height of the box. This ensures the Gaussian distribution’s spread matches the box’s dimensions: $\mathrm{\Sigma }=\left[\begin{array}{cc}{\displaystyle \frac{w^2}{4}}& 0\\ 0& {\displaystyle \frac{h^2}{4}}\end{array}\right]$.

This choice of $\mathrm{\Sigma }$ is geometrically motivated. The semi-axis lengths of the ellipse inscribed within the bounding box are ${\mathrm{\sigma }}_{\mathrm{x}}={\displaystyle \frac{w}{2}}$ and ${\mathrm{\sigma }}_{\mathrm{y}}={\displaystyle \frac{h}{2}}$. For a 2D Gaussian distribution, the ellipse defined by the Mahalanobis distance $\left(\mathrm{x}-\mathrm{\mu }\right)$^⊺ ${\mathrm{\Sigma }}^{-1}\left(\mathrm{x}-\mathrm{\mu }\right)=1$ corresponds to the inscribed ellipse when $\mathrm{\Sigma }=diag\left({\sigma }_x^2,{\sigma }_y^2\right)$. Thus, the Gaussian distribution accurately captures the spatial extent of the bounding box, with highest probability density at the center and decay towards the edges.

The probability density function of the two-dimensional Gaussian distribution is formalized in Eq. (15):

(15) $$f(x|\mu ,\mathrm{\Sigma })=\frac{exp\left(-\frac{1}{2}{\left(x-\mu \right)}^T{\mathrm{\Sigma }}^{-1}\left(x-\mu \right)\right)}{2\pi |\sum {|}^{\frac{1}{2}}}$$where x denotes the positional coordinates, $\mu$ and $\mathrm{\Sigma }$ represent the mean vector and covariance matrix, respectively.

When $\left(\mathrm{x}-\mu \right)$^⊺ ${\mathrm{\Sigma }}^{-1}\left(\mathrm{x}-\mathrm{\mu }\right)=1$ the axis-aligned bounding box can be modeled as a two-dimensional Gaussian distribution, as formalized in Eq. (16):

(16) $$\mu =\left[\begin{array}{c}{c}_x\\ c_y\end{array}\right],\mathrm{\Sigma }=\left[\begin{array}{cc}{\displaystyle \frac{w^2}{4}}& 0\\ 0& {\displaystyle \frac{h^2}{4}}\end{array}\right]$$where $\mathrm{w}=2{\mathrm{\sigma }}_{\mathrm{x}},\mathrm{h}=2{\mathrm{\sigma }}_{\mathrm{y}}$.

To illustrate, consider a ground-truth bounding box for a platelet with coordinates $\left(c_x=100,c_y=150,w=6,h=4\right)$. Using Eq. (16), the Gaussian parameters are:

$${\mu }_{gt}=\left[\begin{array}{c}100\\ 150\end{array}\right],{\mathrm{\Sigma }}_{gt}=\left[\begin{array}{cc}9& 0\\ 0& 4\end{array}\right].$$

A predicted bounding box for the same object with coordinates $\left(c_x=102,c_y=148,w=5,h=5\right)$ yields:

$${\mu }_{gt}=\left[\begin{array}{c}102\\ 148\end{array}\right],{\mathrm{\Sigma }}_{gt}=\left[\begin{array}{cc}6.25& 0\\ 0& 6.25\end{array}\right].$$

The NWD between these two Gaussian distributions is then computed using Eqs. (17) and (19).

Normalized Gaussian Wasserstein distance

The Wasserstein distance is employed to compute the similarity between the two-dimensional Gaussian distributions derived from the ground-truth and predicted bounding boxes, as formulated in Eq. (17):

(17) $$W_2^2({\mu }_1,{\mu }_2)={\Vert m_1-m_2\Vert }_2^2+{\Vert {\mathrm{\Sigma }}_1^{\frac{1}{2}}-{\mathrm{\Sigma }}_2^{\frac{1}{2}}\Vert }_F^2.$$

Since these Gaussian distributions are modeled from bounding boxes A and B, Eq. (17) can be further simplified to Eq. (18):

(18) $$W_2^2(A,B)={\Vert ({\left[c{x}_a,c{y}_a,\frac{w_a}{2},\frac{h_a}{2}\right]}^T,{\left[c{x}_b,c{y}_b,\frac{w_b}{2},\frac{h_b}{2}\right]}^T)\Vert }_2^2.$$

Since the traditional IoU operates within the range [0, 1], the Wasserstein distance must also be normalized to align with this scale. The normalization is formulated in Eq. (19):

(19) $$NWD\left(N_a,N_b\right)=exp\left(-{\displaystyle \frac{\sqrt{W_2^2\left(N_a,N_b\right)}}{C}}\right)$$where $C$ is a parameter representing the average absolute sizeof targets in the dataset.

In the integration of NWD loss, we introduce a hybrid weighting parameter $a$ to balance the contributions of the traditional IoU loss and the NWD loss, as detailed in Eq. (20).

(20) $$Los{s}_{total}=\alpha \times Los{s}_{iou}+\left(1-\alpha \right)\times Los{s}_{iou}.$$

Ablation study

To validate the efficacy of the four methodologies proposed in ‘Materials and Methods’, ablation experiments were conducted on our in-house SBSD dataset. Using YOLOv10m as the baseline model, the experimental results are detailed in Table 2. Figure 14 illustrates the training progression and loss variation as each module is incrementally integrated into the model.

As shown in Table 2, the proposed MCS-Net achieves the highest overall performance on the SBSD dataset, attaining an [email protected] of 0.958 ± 0.0022—a statistically significant improvement of 3.9% over the baseline model (0.919 ± 0.0018), as confirmed by a paired t-test (p < 0.001). It also achieves the highest precision (0.915) and recall (0.922) among all configurations. While maintaining stable parameter counts (16.5M → 16.6M) and reducing computational complexity by 9.6% (GFLOPs: 63.4 → 57.3), the model exhibits a slight decline in inference speed, reflecting a trade-off between accuracy and real-time performance.

When individually integrating the four modules (MSCA, CSD, SHA, NWD), all configurations improve [email protected]. Notably, the CSD module significantly reduces model complexity and computational load (15.2M Params, 56.2 GFLOPs) while preserving high inference speed (108.7 FPS), underscoring its pivotal role in lightweight design. (We rigorously define a model as “lightweight” if it simultaneously satisfies three criteria: parameter count (Params) ≤20M, computational complexity (GFLOPs) ≤60, and inference speed (FPS) ≥100 on a high-end GPU, ensuring deployability on resource-constrained clinical hardware). The NWD module enhances localization robustness for tiny targets via Gaussian distribution modeling, achieving [email protected] = 0.938 (+1.7%) without additional computational overhead. The CSD+NWD lightweight combinationachieves [email protected] = 0.946 with 15.2M parametersand 102 FPS, making it ideal for real-time scenarios. Meanwhile, the MSCA+NWD configurationbalances [email protected] = 0.945 and 80.1 FPS, suitable for applications with moderate accuracy demands.

As illustrated in Fig. 14, all five model configurations exhibit convergence toward stability in both [email protected] and [email protected]–0.95 metrics as training epochs progressively increase. The incremental integration of each proposed module consistently enhances [email protected] and [email protected]–0.95. Notably, in Fig. 14C, the introduction of the CSD detection head significantly reduces the total loss, indicating that its dynamic weight allocation strategy effectively promotes balanced detection across diverse cell categories. Furthermore, the NWD module improves the detection performance for small cells (e.g., platelets and fragmented erythrocytes), resulting in a further reduction in total loss. The full model configuration achieves faster convergence and lower total loss values, demonstrating the synergistic benefits of the proposed modules.

This study conducts a comprehensive investigation into the impact of individual algorithmic modules on the detection performance of six critical blood cell categories: RBC, platelets, schistocytes, granulocytes, monocytes, and lymphocytes. Through rigorous experimentation, we validate the efficacy of each module in addressing detection challenges across cells of varying morphological scales (2–50 μm). Module A denotes the CSD module, Module B represents the SHA module, Module C corresponds to the NWD module, and MCS-Net represents the full configuration integrating CSD, SHA, NWD, and MSCA modules.

Table 3 presents a comprehensive evaluation of how the proposed modules individually and synergistically improve detection accuracy for six blood cell categories in the SBSD dataset. Key observations include: The baseline model (YOLOv10m) achieves strong performance for erythrocytes (RBC: 0.989 ± 0.0004) and granulocytes (GRAN: 0.967 ± 0.0033) but struggles with smaller or morphologically complex targets, particularly schistocytes (0.848 ± 0.0075 AP) and monocytes (MONO: 0.896). The CSD module enhances the detection of granulocytes (GRAN: +0.3% to 0.970) and monocytes (MONO: +3.3% to 0.929) by mitigating feature suppression in dense clusters. SHA: Improves schistocyte detection (0.841 vs. baseline 0.834) through focused nuclear-cytoplasmic correlation modeling but slightly reduces platelet (PLT: 0.872) precision due to attention dilution. NWD: Significantly boosts schistocyte AP (+5.5% to 0.895) by addressing positional sensitivity for fragmented cells. MSCA: Strengthens platelet detection (PLT: 0.899 vs. baseline 0.883) through multi-scale feature fusion but marginally degrades lymphocyte (LYM: 0.948) performance. CSD+NWD: Achieves balanced improvements across scales, elevating PLT (0.906), GRAN (0.984), LYM (0.983), and schistocytes (0.882) while maintaining real-time efficiency. SHA+NWD+MSCA: Demonstrates strong synergy for schistocytes (0.896 AP) and monocytes (0.947), validating the complementary roles of attention and distributional loss. Full MCS-Net Integration: The complete model (MSCA+CSD+SHA+NWD) attains peak performance for GRAN (0.987 ± 0.0023), LYM (0.980 ± 0.0035), MONO (0.980 ± 0.0050), and schistocytes (0.901 ± 0.0039), resolving inter-class conflicts through hierarchical feature fusion. While the SHA module alone slightly reduces PLT precision, its integration within the full MCS-Net framework, complemented by NWD and other components, restores and enhances overall accuracy, achieving a final PLT AP of 0.908 ± 0.0066.

The synergistic value of the full MCS-Net is justified by the complementary insights from Tables 2 and 3. While certain module combinations (e.g., CSD+NWD) achieve high overall mAP and efficiency, they exhibit subtle but important trade-offs in specific cell categories, as detailed in Table 3. The complete MCS-Net architecture is therefore motivated by the need for holistic optimization. By integrating all four components, the model achieves synergistic compensation—where the MSCA provides a multi-scale foundation, the CSD head enables dynamic feature refinement, the SHA enhances large-cell context, and the NWD ensures precise micro-target localization. This results in the most balanced and superior performance across all cell types, particularly for critical pathological targets like schistocytes, without the performance compromises seen in suboptimal combinations.

Figure 15 illustrates a comparative analysis of detection performance between the baseline model and the MCS-Net model on the SBSD dataset. While both models demonstrate robust detection capabilities, subtle differences are observed in their results: Fig. 15A displays three original images, where the black dashed boxes and corresponding enlarged areas represent regions selected for detection comparison. In Fig. 15B, two black arrows indicate undetected platelets. Figure 15B shows that the baseline model misclassifies a schistocyte as an erythrocyte. Figure 15B reveals false positives for schistocytes generated by the baseline model. In contrast, Figure 15C demonstrates that MCS-Net accurately identifies platelets within the blue dashed box. Figure 15C confirms that MCS-Net eliminates such misclassifications, highlighting its superior specificity.

To validate detection efficacy, this study employs XGrad-CAM for visualization analysis of blood smear images (Fu et al., 2020). As shown in Fig. 16A displays two original images. Figure 16B presents the XGrad-CAM visualizations generated by the baseline method. Figure 16C illustrates the XGrad-CAM visualizations produced by the proposed MCS-Net model on the same images. By comparing the visualization results, it is evident that the MCS-Net model demonstrates superior feature extraction capabilities across diverse cell categories and avoids potential missed detections of marginal cells compared to the baseline. Particularly for small cells such as platelets and schistocytes, the heatmaps generated by MCS-Net are more concentrated. These improvements stem from the MSCA module, which efficiently integrates local detail and global semantic features, and the CSD module, whose dynamic weight allocation strategy excels in recognizing cells of varying scales.

Comparative experiments

To validate the superiority of the final model, we conducted a series of comparative experiments against state-of-the-art object detection algorithms (e.g., SSD (Liu et al., 2015), Faster R-CNN, CenterNet (Duan et al., 2019), YOLOv5, YOLOv8), with the results benchmarked in Table 4.

As evidenced by Table 4, the MCS-Net model demonstrates significant comprehensive advantages in blood smear cell detection tasks. It achieves an [email protected] of 0.958 ± 0.0022, marking a 3.9 ± 0.39% improvement over the baseline model YOLOv10 (0.919 ± 0.0018) and substantially outperforming other mainstream algorithms (e.g., Faster R-CNN: 0.831; YOLOv8: 0.871). RT-DETR is an advanced real-time object detector based on the Transformer architecture, whose detection performance surpasses that of most general-purpose object detection models. However, it suffers from limited real-time deployment capability due to its relatively low FPS. Both IHA-YOLO and ASF-YOLO are models specifically designed for cell detection, and they exhibit strong small object detection capabilities. On our SBSD dataset, both achieve [email protected] above 0.9. Although their FPS is superior to that of MCS-Net, they come with higher model complexity. MCS-Net also excels in lightweight design: its parameter count (16.6M) remains comparable to YOLOv10 (16.5M), while reducing computational complexity by 9.6% (57.3 vs 63.4 GFLOPs), surpassing other YOLO variants (e.g., YOLOv8: 79.7 GFLOPs). Although its inference speed (64.9 FPS) is slightly lower than YOLOv10 (107.5 FPS), its accuracy-speed trade-off significantly outperforms two-stage models (e.g., Faster R-CNN: 42.4 FPS). Notably, the lightweight configuration (CSD+NWD) maintains [email protected] = 0.946 at 102 FPS, offering an efficient solution for real-time medical imaging analysis. By integrating multi-scale dynamic interactions, optimized attention mechanisms, and Gaussian distribution modeling, MCS-Net effectively addresses localization deviations for tiny targets, achieving a schistocyte detection precision of 0.901 ± 0.0039. This advancement establishes MCS-Net as a high-precision, robust tool for early screening of DIC and MAHA.

Figure 17 resents a comprehensive comparison of object detection models across three critical metrics: computational efficiency (A), parameter efficiency (B), and accuracy trade-offs (C). Figure 17A (Param vs. GFLOPs) highlights MCS-Net (yellow diamond) as a computationally efficient model, achieving competitive GFLOPs (57.3) with moderate parameters (16.6M), outperforming heavier architectures like YOLOv5 (79.7 GFLOPs) and Faster R-CNN. Figure 17B (Param vs. [email protected]) demonstrates MCS-Net’s superior parameter efficiency, attaining the highest [email protected] among models with comparable parameters. Figure 17C (GFLOPs vs. [email protected]) underscores its optimal balance between speed and accuracy, positioning it closer to the Pareto frontier compared to computationally intensive models (e.g., Faster R-CNN: 42.4 FPS, 0.831 mAP).

Moreover, this study validates the robustness of MCS-Net across multiple datasets, as detailed in Table 5. Specifically, comparative experiments were conducted on three datasets: BCCD, CBC, and SBSD (Alam & Islam, 2019; Xia et al., 2020; Shakarami et al., 2021; Liu, Li & Huang, 2021; Xu et al., 2022; Chen, Tsai & Ho, 2022; Kang et al., 2024b; Mao et al., 2024; Chen & Lu, 2025). Cross-domain evaluations on these datasets demonstrate the model’s exceptional generalization capability and superiority in pathological cell detection: On the general-purpose BCCD dataset, MCS-Net achieves an [email protected] of 0.939 ± 0.0036, representing a 4.7 ± 0.36% improvement over the YOLOv10 model, with PLT detection AP surpassing 0.949 ± 0.0056. On the CBC dataset, MCS-Net maintains its lead with an [email protected] of 0.919 ± 0.0061, achieving an RBC detection precision of 0.874 ± 0.0076 (+3.5 ± 0.76% over the baseline). On the in-house SBSD dataset, MCS-Net comprehensively outperforms the baseline with an [email protected] of 0.958 ± 0.0022, particularly excelling in the detection of schistocytes—a critical pathological indicator—with a precision of 90.1 ± 0.39%, underscoring its dominance in complex morphological cell detection. These results confirm that MCS-Net not only overcomes the limitations of traditional models in detecting pathological cell subtypes but also exhibits stable robustness across diverse data distributions.

Our comparative analysis with prior blood cell detection studies reveals that most existing works were evaluated on the BCCD dataset. The proposed model achieved state-of-the-art performance with an [email protected] of 0.901 ± 0.0039, demonstrating significant superiority in detecting small-scale targets such as PLT. While the model did not achieve the highest accuracy in RBC detection, it exhibited the most robust comprehensive performance overall, balancing precision across multiple cell categories and scales.

Figures 18, 14 illustrates the training convergence of MCS-Net vs the baseline model across the BCCD and CBC datasets, evaluating three metrics: [email protected], [email protected]–0.95, and total loss. Both models achieve near-optimal [email protected] values (~1.0) on BCCD and CBC by epoch 100, demonstrating comparable high-confidence detection accuracy. However, MCS-Net exhibits superior robustness in [email protected]–0.95, reflecting enhanced adaptability to localization variability critical for irregular cell detection (e.g., schistocytes). Notably, MCS-Net achieves significantly lower total loss values with smoother convergence curves, attributed to its MSCA and NWD that mitigate gradient oscillations. The consistent blue (baseline) and orange (MCS-Net) trends across both datasets underscore the model’s cross-dataset reliability, positioning it as a clinically viable solution for precise and stable hematological analysis.

Figures 19 and 20 illustrate the comparative detection performance between the baseline model and MCS-Net on the BCCD and CBC datasets, respectively. The baseline model exhibits missed detections of erythrocytes (red blood cells), whereas MCS-Net significantly reduces such omissions. Additionally, the baseline model suffers from overlapping bounding boxes and localization inaccuracies during detection. The results demonstrate that MCS-Net achieves superior detection capability compared to the baseline, delivering enhanced accuracy and specificity across diverse cell subtypes. Notably, these performance improvements are not confined to the proprietary SBSD dataset but are consistently observed on public benchmarks (BCCD and CBC), underscoring the model’s robust generalization capability.