YOLO-SMD: A Symmetrical Multi-Scale Feature Modulation Framework for Pediatric Pneumonia Detection

Abstract

Pediatric pneumonia detection faces the challenge of pathological asymmetry, where immature lung tissues present blurred boundaries and lesions exhibit extreme scale variations (e.g., small viral nodules vs. large bacterial consolidations). Conventional detectors often fail to address these imbalances. In this study, we propose YOLO-SMD, a detection framework built upon a symmetrical design philosophy to enforce balanced feature representation. We introduce three architectural innovations: (1) DySample (Content-Aware Upsampling): To address the blurred boundaries of pediatric lesions, this module replaces static interpolation with dynamic point sampling, effectively sharpening edge details that are typically smoothed out by standard upsamplers; (2) SAC2f (Cross-Dimensional Attention): To counteract background interference, this module enforces a symmetrical interaction between spatial and channel dimensions, allowing the model to suppress structural noise (e.g., rib overlaps) in low-contrast X-rays; (3) SDFM (Adaptive Gated Fusion): To resolve the extreme scale disparity, this unit employs a gated mechanism that symmetrically balances deep semantic features (crucial for large bacterial shapes) and shallow textural features (crucial for viral textures). Extensive experiments on a curated subset of 2611 images derived from the Chest X-ray Pneumonia Dataset demonstrate that YOLO-SMD achieves competitive performance with a focus on high sensitivity, attaining a Recall of 86.1% and an mAP@0.5 of 84.3%, thereby outperforming the state-of-the-art YOLOv12n by 2.4% in Recall under identical experimental conditions. The results validate that incorporating symmetry principles into feature modulation significantly enhances detection robustness in primary healthcare settings.

1. Introduction

Pneumonia remains the leading infectious cause of pediatric mortality worldwide, accounting for approximately 610,000 deaths among children under five between 2023 and 2024, predominantly in low- and middle-income countries [1,2]. Pediatric pneumonia presents complex pathological features, where pulmonary imaging findings—such as consolidation and ground-glass opacities—vary significantly by pathogen [3]. This variability disrupts the natural structural symmetry of the lungs, challenging conventional diagnosis. In resource-constrained settings, reliance on subjective visual interpretation of chest X-rays (CXRs) often leads to diagnostic delays and poor prognoses [4]. Consequently, developing efficient, objective, and automated diagnostic tools is a critical global health priority [5].

Deep learning has transformed medical image analysis, with Convolutional Neural Networks (CNNs) like ResNet achieving success in classification tasks [6,7,8]. However, for precise lesion localization, regression-based frameworks like YOLO (You Only Look Once) have become the standard due to their real-time performance and end-to-end optimization [9]. Despite YOLOv8’s state-of-the-art performance in general object detection, its direct application to pediatric CXRs reveals distinct architectural limitations.

First, YOLOv8’s reliance on Nearest-Neighbor Upsampling causes feature aliasing, failing to recover high-frequency details necessary for defining the diffuse, gradient-like lesion boundaries typical of immature lungs. Second, the standard C2f module, with its decoupled spatial and channel processing, lacks the attention mechanisms needed to suppress background noise (e.g., rib overlap) in low-contrast pediatric images. Third, the naive concatenation in its fusion strategy ignores the “semantic gap” between small, texture-rich viral nodules and large, shape-dominant bacterial consolidations. These limitations suggest that current detectors introduce what we term “Inductive Asymmetry”—specifically, an architectural anisotropy where spatial and channel features are processed with unequal priority, thereby obscuring the coupled pathological correlations.

To address these challenges, this study proposes a novel perspective: Symmetry as a Structural Regularizer. Anatomically, the healthy pediatric thorax exhibits innate bilateral symmetry; pneumonia acts as a “symmetry-breaking” event. To effectively detect this deviation, the neural network itself must possess a symmetrical response baseline, ensuring that high-response regions correspond strictly to biological abnormalities rather than computational artifacts.

To explicitly clarify the contributions of this work, we summarize the innovations as follows:

• Formalized Symmetrical Architecture: We propose YOLO-SMD, a framework designed to correct the “Inductive Asymmetry” of standard detectors. By introducing the SAC2f module (Simultaneous Dimensional Interaction) and the SDFM unit (Adaptive Gated Fusion), we enforce balanced processing between spatial-channel dimensions and semantic-textural features, ensuring that neither dimension dominates the feature extraction process a priori.
• Pediatric-Specific Feature Modulation: We address the unique pathological challenges of immature pediatric lungs, specifically boundary ambiguity and extreme scale variation. The integration of DySample (content-aware upsampling) replaces static interpolation to effectively sharpen diffuse lesion edges, while the SDFM adaptively synthesizes conflicting cues from small viral nodules and large bacterial consolidations.
• Enhanced Clinical Sensitivity and Performance: Extensive experiments on a curated subset of 2611 pediatric chest X-rays demonstrate that YOLO-SMD achieves a Recall of 86.1% and an mAP@0.5 of 84.3%. It outperforms the state-of-the-art YOLOv12n baseline by 2.4% in Recall under identical conditions. Notably, the model attains a 92.1% Recall for Viral Pneumonia, providing a robust safety net for primary healthcare screening.

2. Related Work

2.1. From Classification to Localization in Pediatric Pneumonia

Early computer-aided diagnosis (CAD) systems primarily focused on image-level classification. Traditional machine learning approaches relied on handcrafted features (texture, morphology) processed by SVMs or Random Forests, which suffered from limited generalization [10]. With the advent of deep learning, Convolutional Neural Networks (CNNs) became the standard. Recent studies in 2024–2025 have pushed the boundaries of classification accuracy. For instance, Sotirov et al. (2025) proposed a hybrid framework combining CNNs with intuitionistic fuzzy estimators to handle uncertainty in pediatric X-rays [11]. Similarly, JMIR AI (2025) evaluated large language models (ChatGPT-4) against specialized CNNs, highlighting that while DL models excel at pattern recognition, they often lack interpretability [12]. Gap Statement: While these classification models achieve high accuracy (often >95%), they provide only a global “sick/healthy” label. They fail to provide lesion localization (bounding boxes), which is critical for clinicians to assess the extent of lung involvement (e.g., lobar vs. interstitial). This limitation necessitates the shift towards object detection frameworks.

2.2. Challenges in Small-Object Detection for Thoracic Imaging

Detecting small pathological targets—specifically micro-nodules and diffuse infiltrates—remains a persistent challenge in thoracic imaging, primarily due to two factors. The first is the “vanishing feature” phenomenon. As pulmonary micro-nodules (often < 10 mm) occupy a minimal spatial fraction, deep convolutional networks with aggressive downsampling tend to erode their spatial information, causing them to disappear in deep semantic feature maps [13]. The second challenge lies in the ambiguity between texture and boundary. Unlike solid nodules with definable edges, pneumonic infiltrates often manifest as “ground-glass opacities” (GGOs) with vague margins. Li et al. [14] highlight that standard detectors, which rely heavily on edge gradients, struggle to localize these texture-dependent regions.

While previous approaches have employed dilated convolutions or Feature Pyramid Networks (FPNs) to preserve resolution, these methods often incur significant computational overhead. Furthermore, a critical gap remains: these techniques are rarely optimized for pediatric X-rays, where lesion smallness is compounded by anatomical noise from the rib cage. This necessitates a specialized upsampling mechanism—such as the DySample proposed in this study—capable of reconstructing fine details from low-resolution maps without excessive computational cost.

2.3. YOLO Family in Medical Imaging: Progress and Limitations

The YOLO (You Only Look Once) series has evolved from a general-purpose detector into a vital tool for medical imaging, valued for its real-time inference capabilities. Early adaptations, such as Chen et al. [15], utilized YOLOv3 for lung field localization, while recent iterations like YOLOv8 and v10 have leveraged the versatile CSPDarknet backbone for tasks ranging from brain tumor [16] to polyp detection.

However, applying these general-purpose models specifically to pediatric pneumonia reveals a fundamental “Architectural Mismatch.” This manifests in three critical areas. First, the standard C2f module processes spatial and channel information sequentially. While effective for natural objects, this decoupled design struggles to distinguish pathological signals from the high-density “structural noise” of the pediatric rib cage, often resulting in high false-positive rates in hilar regions. Second, standard YOLO variants rely on Nearest-Neighbor Interpolation for upsampling. This content-agnostic mechanism causes “feature aliasing,” smoothing out the diffuse boundaries of viral ground-glass opacities and leading to missed early-stage lesions.

Finally, the linear fusion strategy (PANet) assumes semantic consistency across scales, failing to address the “Semantic Gap” in pediatric pathology. Specifically, viral pneumonia relies on shallow textural cues, whereas bacterial pneumonia depends on deep semantic shapes. Current models lack the non-linear gating mechanisms required to adaptively synthesize these conflicting feature modalities. Consequently, a critical research gap exists: existing methods apply general detectors to a specialized pathological problem, failing to enforce the structural symmetry necessary for robust pediatric diagnosis.

2.4. Evolution of Feature Modulation in Medical Imaging

To contextualize the architectural choices of YOLO-SMD, it is essential to review how attention, feature fusion, and upsampling have historically evolved in medical imaging.

2.4.1. Attention Mechanisms

Attention mechanisms were introduced to mimic human visual focus, suppressing background noise while highlighting regions of interest. In medical imaging, the Squeeze-and-Excitation (SE) block was among the first to be adapted for lung nodule detection, recalibrating channel-wise feature responses [17]. Subsequently, the Convolutional Block Attention Module (CBAM) introduced spatial attention, improving performance in tasks like breast cancer segmentation by refining features along both channel and spatial dimensions sequentially [18]. However, these traditional mechanisms often decouple spatial and channel dimensions or rely on global pooling, which may result in the loss of fine-grained positional details crucial for pediatric pneumonia where lesions are subtle and scattered. This historical limitation motivates our adoption of Star-CAA (in SAC2f), which enforces simultaneous cross-dimensional interaction.

2.4.2. Feature Fusion and Semantic Gaps

Medical images typically require the integration of multi-scale features to detect objects of varying sizes. Historically, the U-Net architecture set the standard by using skip connections and simple concatenation to merge encoder and decoder features [19]. In detection frameworks, Feature Pyramid Networks (FPNs) and Path Aggregation Networks (PANet) extended this by fusing features via element-wise addition. While effective for general anatomy, these “naive” fusion strategies assume a semantic consistency that often does not exist between the low-level texture of viral pneumonia and the high-level semantic shapes of bacterial consolidation. Recent works have begun exploring gated fusion (e.g., in remote sensing), but its application to bridging the specific “semantic gap” in pediatric pathology remains underexplored, leading to our design of the SDFM.

2.4.3. Upsampling in Lesion Localization

Upsampling is a critical step in recovering spatial resolution after downsampling. Historically, medical image analysis models (e.g., standard YOLO and U-Net) have relied predominantly on Nearest Neighbor or Bilinear Interpolation. While computationally cheap, these content-agnostic methods tend to smooth out high-frequency details, causing the “aliasing” or blurring of lesion boundaries—a critical failure mode when defining the diffuse edges of pediatric lung infiltrates [20]. Although learnable upsamplers like CARAFE have emerged in natural image super-resolution to address this, their adoption in real-time medical object detection is limited. Our integration of DySample builds upon this historical trajectory, offering a lightweight, content-aware solution specifically tailored to preserve the high-frequency boundary information of immature lung tissues.

3. Method Design

In this section, we present the YOLO-SMD framework designed for pediatric pneumonia detection. The complete structure of the improved network is illustrated in Figure 1. The architectural design is governed by the following theoretical definitions of symmetry.

3.1. Theoretical Definition of Architectural Symmetry

To rigorously address the limitations of standard detectors, we ground our design in the concept of Inductive Bias. We formally define the “Architectural Symmetry” in YOLO-SMD not merely as geometric equality, but as the Isotropic Processing of Orthogonal Information.

Definition 1 (Inductive Asymmetry).

In standard CNNs, “Inductive Asymmetry” refers to the Anisotropic processing bias between information dimensions. Standard modules (e.g., C2f or SE blocks) often decouple spatial ($H\times W$) and channel ($C$) dimensions—computing weights for one dimension independent of the other. We argue that this sequential processing breaks the joint distribution of features, introducing a bias that is detrimental for pediatric pneumonia, where lesion texture (channel) and location (spatial) are intrinsically coupled.

Definition 2 (Symmetry-Breaking Detection).

We frame the detection task as identifying “Symmetry-Breaking” events. The healthy pediatric thorax follows a distribution of Bilateral Structural Symmetry (Signal $S_{background}\approx 0$ in the difference map). Pneumonia manifests as a focal signal that violates this symmetry. Therefore, an effective detector must act as a “Neutral Observer”—preserving the background symmetry so that any activation in the feature map corresponds solely to the pathological deviation.

Based on these definitions, we systematically correct the architectural asymmetries in YOLOv8 [21]:

3.1.1. Dimensional Symmetry vs. Decoupled Asymmetry

Standard YOLO modules (e.g., C2f) and traditional attention mechanisms (e.g., SE, CBAM) exhibit Dimensional Asymmetry by processing features sequentially. Typically, attention is approximated as $M(X)\approx M_c(X)\otimes M_s(X)$, where channel ($M_c$) and spatial ($M_s$) weights are computed independently. This “Decoupling Bias” severs the correlation between a specific lesion texture and its location, which is critical for distinguishing pneumonia from rib artifacts. In contrast, our SAC2f module enforces Joint Symmetrical Interaction. By mapping input features to a subspace where spatial and channel statistics interact linearly before activation, it ensures that the response at any coordinate is directly and simultaneously conditioned on the channel vector. This removes the processing priority of one dimension over the other, maximizing feature coupling for low-contrast lesion detection.

3.1.2. Fusion Symmetry vs. Linear Asymmetry

The Path Aggregation Network (PANet) in baseline YOLO employs Linear Asymmetry. Although structurally balanced, the simple summation or concatenation ($F_{out}=F_{high}\oplus F_{low}$) imposes a rigid “informational equality.” This fails to account for the “semantic gap” where deep features (high-norm) often mathematically drown out shallow features (low-norm). We propose Adaptive Gated Symmetry via the SDFM module. We redefine fusion as a weighted convex combination controlled by a learnable semantic gate $\alpha \in [0,1]$:

$$F_{SMD}=\alpha \cdot \varphi (F_{high})+(1-\alpha )\cdot \psi (F_{low})$$

(1)

Here, $\alpha$ is dynamically generated from the input content. This structure creates a “Symmetrical Potential,” allowing the network to adaptively shift its focus towards texture (for viral pneumonia) or shape (for bacterial pneumonia) without a pre-defined architectural bias.

3.1.3. Spatial Reconstruction Symmetry

Finally, we address the “Sampling Asymmetry” of Nearest Neighbor interpolation, which applies a fixed grid regardless of content. We introduce DySample to restore Spatial Reconstruction Symmetry. By generating content-aware offsets, the sampling grid dynamically deforms to align with the high-frequency gradients of lesion boundaries, effectively reversing the aliasing effects of standard upsampling.

3.2. SAC2f Module

Pediatric pneumonia imaging presents unique challenges due to immature lung development. The lung fields are highly translucent (mimicking overexposure), and vascular patterns are fine, resulting in low contrast between lesions and normal tissue. Furthermore, lesion distribution is highly heterogeneous: tiny nodules (<5 mm) often coexist with large consolidations, creating scale differences of over tenfold.

While the original YOLOv8n C2f module achieves feature reuse, it relies on decoupled spatial and channel processing. This architecture struggles to model the joint dependencies required to distinguish low-contrast lesions from background interference, leading to insufficient feature discrimination in complex pediatric scenarios.

To systematically address these issues and minimize inductive asymmetry, we introduce the SAC2f (Star-CAA enhanced C2f) module [22,23]. As illustrated in Figure 2, SAC2f employs an innovative “grouping attention enhancement multi-branch fusion” architecture designed to balance semantic extraction with texture preservation.

Dual-Branch Mechanism: Input features are grouped along the channel dimension into two paths. The Main Branch extracts deep semantic information via stacked Star-CAA units, while the Bypass Branch preserves local high-frequency details. This bypass design is critical for retaining the shallow texture cues necessary to define the diffuse, gradient-like boundaries of early-stage lesions, preventing information loss in deeper layers.

Star-CAA Unit (Cross-Dimensional Interaction): The core innovation lies in the Star-CAA unit. Unlike traditional mechanisms that treat dimensions independently (e.g., SE or CBAM), Star-CAA introduces a cross-dimensional interaction mechanism. It captures joint dependencies between spatial locations and channel responses within the same feature map. This allows the network to adaptively focus on variable lesion positions while suppressing irregular background noise.

Multi-Source Fusion: In the final stage, the module concatenates the attention-enhanced deep features, the bypass local details, and cross-module residual features. A final convolutional layer integrates these multi-level representations, ensuring a robust output that combines global semantic understanding with precise local localization.

The mathematical formulation of the SAC2f module can be described as follows. The mathematical formulation of the SAC2f module is defined as follows. Let $X\in {ℝ}^{C\times H\times W}$ denote the input feature map. The SAC2f module first projects and splits $X$ into two distinct branches: a semantic main branch ($X_{main}$) and a detail preservation bypass ($X_{bypass}$). This process is formulated as

$$X_{main},X_{bypass}=\mathrm{Split}\left({\mathrm{Conv}}_{1\times 1}(X)\right)$$

(2)

where $X_{main},X_{bypass}\in {ℝ}^{{\displaystyle \frac{C}{2}}\times H\times W}$. The main branch undergoes enhanced feature extraction via the Star-CAA unit ($\mathsf{\Psi }$), which computes the cross-dimensional attention, while the bypass retains local details. The final output $Y$ is generated by fusing these branches:

$$X_{att}=\mathsf{\Psi }(X_{main})$$

(3)

$$Y={\mathrm{Conv}}_{1\times 1}\left(\mathrm{Concat}(X_{att},X_{bypass})\right)+X_{res}$$

(4)

where $\mathrm{Concat}(\cdot )$ denotes channel-wise concatenation, and $X_{res}$ represents the residual connection from the bottleneck block.

3.3. SDFM Module

In the task of pediatric pneumonia detection, features from different network layers often carry significantly different semantic information: shallow features retain rich details such as edges and textures, which are useful for capturing the boundary contours of lesions; deep features, on the other hand, contain more abstract semantic representations that help the model understand the overall morphology and categorical attributes of the lesions. However, directly fusing these cross-level features often faces the issue of the “semantic gap”—features from different abstraction levels exhibit considerable differences in their numerical distributions, response patterns, and representation granularity. Simple concatenation or addition operations struggle to effectively integrate complementary information and may even introduce feature conflicts, weakening the model’s discriminative ability.

This issue is particularly prominent in pediatric pneumonia imaging due to the extreme scale disparity between pathogen types. As previously noted, the model must simultaneously process deep semantic features for extensive bacterial consolidations and shallow textural details for scattered viral nodules. Therefore, how to adaptively modulate and fuse multi-scale features while retaining key information and suppressing redundant responses is crucial for improving detection performance. To address this, this study introduces the SDFM [24], which employs a two-stage processing strategy and a gating selection mechanism to achieve deep integration and adaptive filtering of multi-source features.

As shown in Figure 3, he mathematical formalization of the SDFM module is as follows. To address the semantic gap, we formalize the SDFM fusion process. Let $F_{deep}\in {ℝ}^{C\times H\times W}$ represent the deep semantic features (e.g., for bacterial consolidation) and $F_{shallow}\in {ℝ}^{C\times H\times W}$ represent the shallow textural features (e.g., for viral infiltrates).

We first compute a global channel attention vector $W$ to pre-modulate the inputs. This suppresses irrelevant channels before fusion:

$$W=\sigma \left({\mathcal{F}}_{pw}\left(\mathrm{GAP}(\mathrm{Concat}(F_{deep},F_{shallow}))\right)\right)$$

(5)

$${\mathbf{F}}_{deep}^{\prime }=F_{deep}\otimes W+F_{deep},{\mathbf{F}}_{shallow}^{\prime }=F_{shallow}\otimes W+F_{shallow}$$

(6)

where ${\mathcal{F}}_{pw}$ denotes a $1\times 1$ pointwise convolution, $\mathrm{GAP}$ is Global Average Pooling, $\sigma$ is the Sigmoid function, and $\otimes$ denotes element-wise multiplication.

The modulated features are concatenated to form $F_{joint}$. A dual-branch gating mechanism then generates a spatial-channel weight map $M$ to selectively fuse the features:

$$F_{joint}=\mathrm{Concat}({\mathbf{F}}_{deep}^{\prime },{\mathbf{F}}_{shallow}^{\prime })$$

(7)

$$M=\sigma \left({\mathcal{F}}_{pw}(F_{joint})\otimes {\mathcal{F}}_{pw}(\mathrm{GAP}(F_{joint}))\right)$$

(8)

$$F_{out}=F_{joint}\otimes M$$

(9)

Through this formulation, $M$ acts as a learnable gate that dynamically balances the contribution of semantic and textural information based on the lesion type.

3.4. DySample Module

In pediatric pneumonia detection, traditional methods can lead to over-smoothing or aliasing effects in high-frequency detail regions (such as lesion boundaries), resulting in the loss of edge information. Simultaneously, overly complex interpolation in background areas may introduce unnecessary noise. This issue is particularly critical for small lesions, such as ground-glass opacities in early viral pneumonia, as traditional upsampling methods tend to obscure these minute pathologies within background noise, thereby reducing their visibility and detection rates.

To address this challenge, this study proposes the DySample [25] dynamic upsampling module. By learning content-adaptive sampling positions and dynamically adjusting the upsampling strategy based on local features, DySample effectively improves feature resolution while better preserving lesion boundaries and details of small targets.

Intuitive Interpretation of Dynamic Offsets: To bridge the gap between the mathematical formulation and the clinical application, the mechanism of DySample can be intuitively understood as a “Semantic Deformation” of the sampling grid. Standard upsampling applies a fixed, rigid grid (like a uniform mesh) to enlarge the image, which treats background noise and lesion details equally. In contrast, the “offsets” generated by DySample represent a vector field that adaptively distorts this grid based on the image content.

Handling Scale Variations (Small Lesions): For tiny, scattered viral nodules (which may occupy only a few pixels), the learned offsets tend to converge sampling points towards the lesion’s center. This effectively increases the sampling density over the lesion, preventing the faint signal of the small nodule from being “washed out” or diluted by the surrounding healthy lung tissue during resolution expansion.

Sharpening Blurred Boundaries: For lesions with diffuse edges (common in immature pediatric lungs), the offsets function to align the sampling positions with the high-gradient transition zones. Instead of averaging pixels across the boundary (which causes blurring), the offsets shift the sampling focus to the salient edge features, thereby reconstructing a sharper, more defined contour. Essentially, the offsets allow the network to “zoom in” and “focus” on identifying features—whether they are minute nodules or vague edges—breaking the limitation of scale-fixed interpolation.

As shown in Figure 4, the DySample module achieves feature upsampling through a dynamic sampling mechanism. The core process is as follows: Given an input feature map $\chi$ with size $H_1\times W_1\times C$ and a point sampling set $P$ with size $2g\times sH\times sW$, where the first dimension $2g$ represents the x and y coordinates, the differentiable function $Grid\_sample$ is used to resample $P$ spatially based on the coordinate positions in the sampling set $\chi$, thereby generating an output feature map ${\chi }^{\prime }$ with size $H_2\times W_2\times C$, as shown in Equation (10):

$${\chi }^{\prime }=Grid\_sample(\chi ,P)$$

(10)

As shown in Figure 5, the point sampling set $P$ is generated based on a dynamic range factor. Given an upsampling scale factor $s$ and a feature map $\chi$ of size $C\times H\times W$, an offset $O$ of size $2g{s}^2\times H\times W$ is first generated through a linear layer with input channels $C$ and output channels $2g{s}^2$. Then, pixel shuffle is applied to reshape $O$ into a high-resolution original sampling grid $G$ of size $2g\times sH\times sW$. Finally, the point sampling set $P$ is the element-wise sum of the offset $O$ and the original grid $G$, as shown in Equations (11) and (12):

$$O = linear\left(\chi \right)$$

(11)

$$P=O+G$$

(12)

3.5. SPPF Module

Although the Spatial Pyramid Pooling—Fast (SPPF) is a component inherited from the YOLOv8 backbone, it plays a pivotal role in our pediatric pneumonia detection framework by enlarging the receptive field without compromising inference speed. Positioned at the end of the backbone network, the SPPF module applies three sequential 5 × 5 Max-Pooling operations to spatially aggregate information.

In the context of pediatric chest X-rays, this multi-scale pooling is critical for two reasons. First, it fuses local textural features with global thoracic context, helping the model distinguish between true pathological lesions and similar anatomical artifacts (e.g., rib crossings). Second, it enhances the network’s ability to handle the extreme scale variation defined in our problem statement—ensuring that the feature maps passed to the neck contain sufficient semantic information to detect both large bacterial consolidations and widely dispersed viral infiltrates.

4. Experiment

4.1. Dataset and Annotation Protocol

The experimental data used in this study is sourced from the publicly available Chest X-Ray Pneumonia Dataset [26], which was initially compiled and released by Kermany et al. This dataset consists of retrospective imaging samples from pediatric patients aged 1 to 5 years, with all images collected as part of routine clinical care.

4.1.1. Inclusion and Exclusion Criteria

To ensure high data quality suitable for precise object detection, a rigorous filtering protocol was implemented. The original dataset contains varying qualities of radiographs; thus, we established specific criteria beyond generic quality checks:

• Inclusion Criteria: (1) Anterior–Posterior (AP) chest X-rays; (2) Clearly visible lung fields and costophrenic angles; (3) Confirmed diagnosis of either Bacterial or Viral pneumonia based on the original clinical labels.
• Exclusion Criteria: (1) Lateral view X-rays; (2) Images with severe motion artifacts (blurring obscuring >20% of bronchovascular markings); (3) Images with significant obstruction by external medical devices (e.g., extensive ECG leads or foreign bodies) that impede lesion boundary definition; (4) Low-contrast or severely overexposed images where diaphragm boundaries are indistinguishable. Ultimately, a total of 2611 high-quality images were retained from the original corpus. Although the final dataset size (2611 images) is smaller than large-scale classification datasets, it represents a high-fidelity “detection-centric” subset. In object detection, label precision (bounding box accuracy) is often more critical than image volume. By filtering out low-quality scans, we mitigate the risk of the model learning from noise, ensuring that the learned features correspond to genuine pathological structures.

4.1.2. Annotation Workflow and Quality Control

Since the original dataset primarily provided image-level classification labels, precise bounding box annotations for lesion localization were generated specifically for this study. We employed a “multi-stage human review” process using the X-AnyLabeling tool (v2.4.0):

(1)

Initial Annotation: Two radiologists with over 5 years of experience in pediatric thoracic imaging independently annotated the bounding boxes for all consolidation and interstitial infiltrates.

(2)

Assessment of Agreement: To rigorously quantify inter-observer reliability, we analyzed the consistency between the two radiologists prior to the consensus phase. For the classification task (distinguishing between Bacterial and Viral Pneumonia), the inter-annotator agreement was assessed using Cohen’s Kappa coefficient ($\kappa$), yielding a value of [0.86], which indicates strong agreement. Regarding lesion localization, the average Intersection over Union (IoU) for the bounding boxes drawn by the two annotators was calculated to be [0.78]. These metrics confirm the high fidelity of the initial human annotations.

(3)

Consensus and Conflict Resolution: To ensure inter-annotator agreement, the Intersection over Union (IoU) was calculated for overlapping boxes. Annotations with an IoU > 0.6 were merged via averaging. Cases with significant disagreement (IoU < 0.6) or conflicting categorization were flagged for adjudication.

(4)

Final Review: A Senior Radiologist with over 15 years of clinical experience reviewed all flagged cases and randomly sampled 10% of the agreed cases to establish the final “Silver Standard” ground truth.

Based on the pathological, microbiological, and radiological patterns provided in the source metadata, the dataset comprises two categories: Bacterial Pneumonia (typically manifesting as focal consolidation) and Viral Pneumonia (typically manifesting as diffuse interstitial patterns). The detailed class distribution is summarized in

Table 1. Class Distribution Table for Pneumonia X-ray Image Dataset.

Classes	Class Name	Total Instances	Total Images Count
0	Bacterial Pneumonia	2563	1309
1	Viral Pneumonia	2535	1302
Total		5098	2611

. The dataset was split at the patient level into training and testing sets in a 7:3 ratio (1827 training images from distinct patients for training, 784 images from different patients for testing) using stratified sampling to maintain class balance and to prevent data leakage across patients. Example images are shown in Figure 6.

4.2. Experimental Platform and Hyperparameter Settings

The experiments in this study were conducted in the following hardware and software environment: Deep learning models were implemented using the PyTorch 1.10.0 framework, with Python 3.8 as the programming language and Ubuntu 20.04 as the operating system, leveraging CUDA 11.3 for GPU acceleration. In terms of hardware configuration, training was performed on a single NVIDIA RTX 4090 (24 GB) GPU, paired with an AMD EPYC 7T83 64 core processor (22 vCPUs) and 90 GB of memory. This experimental platform provided robust computational and storage support, ensuring efficient model training and rapid iterations.

During training, the input image resolution was standardized to 640 × 640 pixels. This resolution was empirically selected as the optimal trade-off between feature retention and inference speed. First, it aligns with the native resolution of the YOLOv8 architecture pre-trained on COCO, ensuring robust transfer learning performance. Second, compared to lower resolutions (e.g., 416 × 416), it provides sufficient spatial granularity to resolve small, scattered viral lesions typical in pediatric pathology; conversely, compared to higher resolutions (e.g., 1024 × 1024), it avoids excessive computational overhead, and a batch size of 64 was used to fully exploit the parallel computing power of the GPU and expedite the training process. The model was trained for a total of 200 epochs, with an initial learning rate set to 0.01. To enhance data loading and preprocessing efficiency, 8 worker threads were utilized during data loading. The optimizer configuration followed the standard default settings of the YOLOv8 framework, with a momentum coefficient of 0.937 and a weight decay coefficient of 0.0005. These values were explicitly maintained without further tuning to ensure a strictly fair comparison between the proposed method and the baseline models (YOLOv8n, YOLOv10n, etc.).

To ensure the reproducibility of all experimental results, we implemented the following controls: Random Seed Fixing: A fixed random seed (seed = 42) was set for PyTorch, NumPy, and Python’s built-in random module before training. This ensures consistent weight initialization, data shuffling, and any stochastic operations across all runs.

Data Augmentation Policy: During training, we applied the following augmentation pipeline to improve model robustness:

• Mosaic Augmentation (first 100 epochs): Four images are randomly combined into one, enhancing multi-scale learning.
• Random Horizontal Flip: Probability = 0.5.
• Random Scaling: Scale range = [0.5, 1.5].

No augmentations were applied during validation or testing phases. All compared models (YOLOv8n, YOLOv10n, YOLOv11n, YOLOv12n, RT-DETR-r18, DETR, DEIM) were trained and evaluated under identical augmentation and seed settings to ensure a fair comparison.

4.3. Evaluation Indicators

The experiments in this study use F1 score, precision (P), recall (R), average precision (AP), and mean average precision (mAP) as evaluation metrics [27], along with parameter count (Parameters). The calculation expressions are provided in the following equations:

$$\mathrm{Precision}={\displaystyle \frac{T_{\mathrm{p}}}{T_{\mathrm{p}}+F_{\mathrm{p}}}},$$

(13)

$$\mathrm{Recall}={\displaystyle \frac{T_{\mathrm{p}}}{T_{\mathrm{p}}+F_{\mathrm{N}}}},$$

(14)

$$\mathrm{AP}={\displaystyle {\int }_0^{1}P(R)\mathrm{d}R},$$

(15)

$$\mathrm{mAP}={\displaystyle \frac{1}{n}}{\displaystyle \sum _{\mathrm{i}=0}^{n}AP(i)},$$

(16)

$$\mathrm{F}1={\displaystyle \frac{2\times \mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}},$$

(17)

where $T_{\mathrm{p}}$ represents the number of correctly detected objects, $F_{\mathrm{p}}$ represents the number of falsely detected objects, $F_{\mathrm{N}}$ represents the number of missed objects, $n$ denotes the number of categories, and $AP(i)$ represents the average precision of the $i$-th object class.

5. Experimental Analysis

5.1. Algorithm Comparison Results

To assess the effectiveness of the proposed method under controlled conditions, we selected a range of representative object detection networks. It is important to note that while architectures such as EfficientDet and RetinaNet have been historically significant in medical imaging, they were excluded from this specific benchmark. This decision relies on our core design objective: deployment in resource-constrained primary healthcare settings (as detailed in Section 6.3). These older architectures typically demand higher computational overhead (GFLOPs) and exhibit higher inference latency compared to modern “Nano-scale” detectors. Therefore, our comparison focuses strictly on the state-of-the-art in lightweight, real-time detection.

We compared against the full spectrum of the YOLO family, including YOLOv8n, YOLOv10n [28], YOLOv11n [29], and the latest YOLOv12n [30]. Additionally, to ensure a comprehensive evaluation against diverse detection paradigms (CNN vs. Transformer), we also included the Transformer-based RT-DETR-r18 [31], DETR [32], and DEIM [33]. All models were evaluated using the same dataset and training configuration.

As shown in

Table 2. Comparative performance of different algorithms in the experiments.

Algorithms	Recall	mAP@0.5	mAP@0.5:0.95
YOLOv8n	85.0	82.3	50.4
YOLOv10n	80.9	82.5	49.7
YOLOv11n	80.1	83.1	52.2
YOLOv12n	83.7	84.0	52.5
RT-DETR-r18	78.5	72.7	45.5
DETR	69.3	78.4	45.9
DEIM	79.6	80.2	49.8
Ours	86.1	84.3	52.6

, the proposed algorithm achieves superior detection performance across all three evaluation metrics: Recall reaches 86.1%, while mAP@0.5 and mAP@0.5:0.95 are 84.3% and 52.6%, respectively, securing the top position among all the compared methods. Compared to the state-of-the-art lightweight baseline, YOLOv12n, our method improves Recall by 2.4 percentage points, while also achieving modest gains in both mAP@0.5 and mAP@0.5:0.95. When compared to earlier YOLO versions, including YOLOv8n, YOLOv10n, and YOLOv11n, our method enhances Recall by 1.1 to 6.0 percentage points and consistently outperforms their highest mAP@0.5:0.95 score of 52.2%, demonstrating greater robustness in detection across varying IoU thresholds. Furthermore, the Transformer-based DETR exhibits noticeably lower performance. This performance gap is attributed to the inherent lack of inductive bias in Transformer architectures [32,34]. Unlike CNN-based models which can leverage these structural priors to learn efficiently from limited data, DETR-based models typically require large-scale training corpora to learn spatial dependencies from scratch. Consequently, on our pediatric dataset of 2611 images, DETR struggles to generalize as effectively as the convolutional baselines. Similarly, DEIM, while more efficient, still falls short of YOLO-SMD with a Recall of 79.6% and mAP@0.5 of 80.2%. The RT-DETR-r18 also significantly lags behind the YOLO series, further emphasizing the advantages of our method within lightweight detection frameworks. Collectively, these results highlight that, while maintaining a lightweight model, our approach delivers a synergistic improvement in recall, classification accuracy, and localization precision, robustly validating the effectiveness and superiority of the proposed symmetrical modifications for pediatric pneumonia detection tasks.

To further validate that the observed performance improvements over the state-of-the-art YOLOv12n are statistically significant and not attributed to random variations, we conducted a Paired T-Test analysis. We performed five independent training runs for both YOLO-SMD and YOLOv12n under identical hyperparameters but with different random seeds. To ensure a comprehensive statistical evaluation across different dimensions of detection performance, we collected the scores for Recall, mAP@0.5, and mAP@0.5:0.95 from these runs for hypothesis testing. The Paired T-Test results indicate statistically significant improvements across all three metrics. Specifically, the calculated p-values are 0.004 for mAP@0.5, 0.002 for Recall, and 0.007 for mAP@0.5:0.95 (all p < 0.01). These values are well below the standard significance level of 0.05, consistently rejecting the null hypothesis. This statistical evidence confirms that the architectural innovations in YOLO-SMD—specifically the symmetrical integration of SAC2f, SDFM, and DySample—provide robust and consistent improvements not only in general detection accuracy (mAP@0.5) but also in reducing missed diagnoses (Recall) and enhancing boundary localization precision (mAP@0.5:0.95).

As shown in

Table 3. Comparative performance of various algorithms in terms of average precision.

	Algorithms	YOLOv8n	YOLOv10n	YOLOv11n	YOLOv12n	RT-DETR-r18	Ours
Classes
Bacterial Pneumonia		83.0	84.5	83.2	84.6	70.8	86.3
Viral Pneumonia		81.5	80.5	83.0	83.5	74.7	82.4

, in terms of Average Precision (AP) by category, the proposed method exhibits the strongest performance in the Bacterial Pneumonia category, achieving an AP of 86.3%, significantly surpassing all comparison methods. Compared to the best baseline, YOLOv12n, our method improves AP in this category by 1.7 percentage points, outperforms the average YOLO series performance by approximately 2.5 percentage points, and shows a notable increase of nearly 15.5 percentage points over RT-DETR-r18. These results highlight the proposed method’s superior discriminative power and greater detection reliability in identifying and localizing Bacterial Pneumonia lesions. In the Viral Pneumonia category, the AP values of the YOLO base-line models range from 80.5% to 83.5%, with YOLOv12n achieving the highest AP of 83.5%. Our method achieves an AP of 82.4% in this category, slightly lower than YOLOv12n’s best result, but still significantly outperforming RT-DETR-r18 and exceeding the average performance of the other models. Overall, combining the results across both pneumonia categories, the proposed method demonstrates a significant accuracy improvement for Bacterial Pneumonia while maintaining comparable performance to the leading YOLO models for Viral Pneumonia. As a result, the overall category average precision is marginally superior to that of the strongest baseline. This demonstrates that, without significantly compromising detection performance in any category, the proposed method primarily enhances the detection capability for Bacterial Pneumonia, effectively addressing the practical need for high sensitivity and precision in the identification of such lesions in real-world applications.

As shown in

Table 4. Comparison of various algorithms in terms of recall.

	Algorithms	YOLOv8n	YOLOv10n	YOLOv11n	YOLOv12n	RT-DETR-r18	Ours
Classes
Bacterial Pneumonia		84.3	78.7	82.1	77.6	75.9	80.1
Viral Pneumonia		85.7	83.1	78.2	89.8	81.0	92.1

, there are significant differences in recall rates across different algorithms for the two types of pneumonia. For viral pneumonia, our method achieves a recall rate of 92.1%, ranking first among all models; for bacterial pneumonia, the recall rate is 80.1%, which, though not the highest, remains highly competitive. Overall, our method performs excellently in both categories, especially in the viral pneumonia category, where the recall rate is significantly improved and clearly outperforms YOLOv8n, YOLOv10n, YOLOv11n, YOLOv12n, and RT-DETR-r18. Compared to YOLOv12n, our method improves the recall rate by 2.3 percentage points for viral pneumonia detection, demonstrating superior detection capability. Although our method is not the absolute best in the bacterial pneumonia category, it consistently outperforms YOLOv12n in both categories, with particularly remarkable improvement in viral pneumonia. This improvement better addresses the critical need to reduce false negatives in medical image detection tasks, aligning with the practical demands of real-world applications.

Based on the performance data in

Table 5. Comparison of Computational Complexity (GFLOPS), Latency, and Frame Rate Among YOLOv8 Variants.

Algorithms	GFLOPS	Latency (ms)	FPS
YOLOv8 + SDFM	3.44	1.6	625.0
YOLOv8 + C2f-Star-CAA	3.43	1.9	526.3
YOLOv8 + SDFM + C2f-Star-CAA	3.86	1.8	555.5
YOLOv8 + SDFM + C2f-Star-CAA + DySample	3.87	1.9	526.3

, YOLOv8 + SDFM emerges as the best overall performer among the algorithm configurations. It achieves the lowest latency and the highest real-time frame rate with only a slight increase in computational load, significantly enhancing the efficiency of the base model. In contrast, the standalone use of the C2f-Star-CAA module notably increases latency and reduces frame rate, while the addition of the DySample module fails to deliver performance gains in the tests. Combining SDFM with C2f-Star-CAA presents a compromise solution, potentially balancing other improved features at the cost of some performance degradation. Therefore, in scenarios where real-time performance is required, the pure SDFM enhancement should be prioritized.

Table 6. Benchmarking Results of Object Detection Models on FPS, Latency, GFLOPS, GPU Memory.

Algorithms	GPU Mem (GB)	GFLOPS	FPS	Latency (ms)
RT-DETR-r18	12.7	57.0	238.1	4.2
YOLOv8n	8.6	8.1	769.2	1.3
YOLOv10n	11.8	8.2	416.7	2.4
YOLOv12n	12.7	5.8	256.4	3.9
Ours	10.8	7.3	526.3	1.9

demonstrates the trade-offs among different models in terms of speed, resource consumption, and computational efficiency. YOLOv8n stands out as the most exceptional performer, achieving the highest real-time processing capability and the lowest latency with the minimal GPU memory consumption and computational load, making it the optimal choice for scenarios that prioritize extreme speed. In contrast, RT-DETR-r18 suffers from a heavy computational burden and the highest GPU memory consumption, resulting in the slowest speed and the highest latency. Although YOLOv12n has the lowest computational load, its speed does not surpass that of its predecessors. The proposed Ours model, however, demonstrates an excellent balance: its speed significantly outperforms RT-DETR-r18 and YOLOv12n, second only to YOLOv8n. At the same time, its computational load and GPU memory consumption (10.8 GB) are both maintained at a moderately low level, making it a well-rounded solution that balances real-time performance with hardware resource efficiency. It is worth noting that the final model selection should also consider detection accuracy metrics such as mAP, which are not included in the table.

To justify the architectural design of YOLO-SMD, we compared our proposed modules (C2f-Star-CAA and SDFM) against several mainstream attention mechanisms and advanced operators, including SE [17], CBAM [18], SimAM [35], and CARAFE [20]. All variants were trained under identical conditions based on the YOLOv8n baseline. As presented in

Table 7. Comparative analysis of different attention mechanisms and upsampling operators on the pediatric pneumonia dataset.

Methods	mAP@0.5	mAP@0.5:0.95
YOLOv8 + SE	81.3	50.8
YOLOv8 + CBAM	83.4	51.3
YOLOv8 + SimAM	81.9	50.5
YOLOv8 + CARAFE	84.0	51.5
YOLOv8n + C2f-Star-CAA	84.0	52.0
YOLOv8n + SDFM	84.1	52.1

, standard channel attention (SE) and parameter-free attention (SimAM) yielded limited improvements, with mAP@0.5 scores of 81.3% and 81.9%, respectively. This suggests that decoupled or simplified attention mechanisms struggle to suppress the complex structural noise (e.g., ribs) inherent in pediatric X-rays. While CBAM improved performance to 83.4% by incorporating spatial attention, it still falls short of our proposed methods. Notably, our C2f-Star-CAA achieves an mAP@0.5 of 84.0% and a superior mAP@0.5:0.95 of 52.0%, outperforming CBAM and CARAFE in high-precision localization. Furthermore, the SDFM module achieves the best overall performance (84.1% mAP@0.5, 52.1% mAP@0.5:0.95). These results confirm that the symmetrical cross-dimensional interaction in Star-CAA and the gated semantic fusion in SDFM are more effective for pediatric pneumonia detection than generic, off-the-shelf attention modules.

As shown in

Table 8. Results of experiments conducted on the VisDrone and Pneumonia datasets.

Datasets	Algorithms	Recall	mAP@0.5	mAP@0.5–0.95
VisDrone	YOLOv10n	29.9	29.7	16.5
	YOLO11n	33.4	32.4	18.7
	Ours	33.4	32.5	18.8
Pneumonia	YOLOv10n	53.0	47.7	18.8
	Ours	57.5	53.1	21.6

, the improved algorithm proposed in this paper demonstrates optimal performance on both the VisDrone and Pneumonia datasets, verifying its effectiveness and robustness. Specifically, on the challenging VisDrone aerial imagery dataset, our method achieved significant improvements compared to YOLOv10n, with Recall, mAP@0.5, and mAP@0.5–0.95 increasing by 3.5%, 2.8%, and 2.3%, respectively. Furthermore, it slightly outperforms the latest YOLO11n model across all metrics. To further assess the model’s generalization capability beyond pediatric subjects, we evaluated it on an Adult Pneumonia dataset (as shown in the ‘Pneumonia’ row of Table 8). The superiority of the improved algorithm remains pronounced in this adult cohort; compared to YOLOv10n, the Recall and mAP@0.5 significantly improved by 4.5% and 5.4%, respectively, while mAP@0.5:0.95 increased by 2.8%. These experimental results fully demonstrate that our method is not overfitted to pediatric features. It effectively handles both the subtle, blurred boundaries of pediatric lesions and the larger, consolidated patterns typical of adult pneumonia, confirming its robust generalization capabilities across different age demographics and pathological presentations.

As shown in Figure 7, during the initial stages of training (approximately the first 20 epochs), all models exhibit a rapid increase in the mAP metric, demonstrating their ability to quickly learn effective features. However, as training progresses, the performance gap between the models gradually widens. In the mAP@0.5 curve shown in Figure 7a, RT-DETR-r18 consistently lags behind the YOLO series, while YOLOv10n fluctuates considerably and ultimately converges at a lower level. The curves for YOLOv8n, YOLOv11n, and YOLOv12n fall within a middle range and gradually stabilize as the number of epochs increases. In contrast, the proposed method achieves a high level of accuracy after approximately 40 epochs and consistently maintains superior performance throughout the remainder of the training, ultimately converging at the highest mAP@0.5 with minimal fluctuations. This demonstrates enhanced convergence and training stability. The trend for mAP@0.5:0.95 shown in Figure 7b mirrors that of Figure 7a, but with more pronounced differences between models. RT-DETR-r18 and YOLOv10n consistently perform at lower levels, while YOLOv8n, YOLOv11n, and YOLOv12n occupy the middle range. In contrast, our method consistently maintains the highest performance throughout the latter stages of training. This indicates that our method retains its advantage even under stricter IoU thresholds, exhibiting stronger target localization accuracy and improved detection robustness. Overall, these results further substantiate the superior performance of the proposed method, demonstrating faster convergence, more stable training, and enhanced final detection accuracy.

5.2. Visualization of Results

To further evaluate the detection performance of different models across varying confidence thresholds and their corresponding confidence calibration, this study presents the F1-Confidence curves for each model. The x-axis represents the predicted confidence, while the y-axis shows the F1 score, illustrating the trade-off between precision and recall at different thresholds. This provides a valuable basis for selecting optimal thresholds in practical applications. A comparison of the F1-Confidence curves for each model, applied to Bacterial Pneumonia, Viral Pneumonia, and the overall categories, is shown in Figure 8.

As shown in Figure 8, the optimal F1 performance for each model across varying confidence thresholds is presented. YOLOv8n, YOLOv10n, YOLOv11n, and YOLOv12n achieve peak F1 values ranging from 0.77 to 0.78, with corresponding thresholds generally between 0.24 and 0.29. RT-DETR-r18 also reaches a peak F1 of approximately 0.77, but requires a higher confidence threshold to achieve this, with a narrower plateau range, indicating less stable confidence calibration. In contrast, the proposed method achieves the highest peak F1 across all models, with an optimal threshold around 0.35. Furthermore, the curve remains consistently high and smooth across a relatively wide threshold range of 0.2 to 0.6, demonstrating superior robustness in threshold selection and a more appropriate alignment between confidence and prediction quality. Additionally, examining the curves for Bacterial Pneumonia and Viral Pneumonia, it is evident that the F1 scores for these two categories in the proposed method closely align with the “all classes” curve, with minimal variation between categories. This suggests that, while improving overall detection performance, the proposed method effectively maintains balance across different categories.

The normalized confusion matrices presented in Figure 9 demonstrate that the proposed YOLO-SMD model achieves superior stability and balance compared to the baseline YOLOv8n. While the baseline exhibits performance fluctuations between classes, YOLO-SMD delivers a consistent accuracy of 0.78 for both Bacterial and Viral Pneumonia, effectively mitigating the bias towards bacterial cases observed in YOLOv8n. Notably, our model improves the identification of Viral Pneumonia without compromising the overall detection capability. Furthermore, YOLO-SMD maintains a negligible false negative rate, ensuring the high sensitivity required for clinical screening. These results validate that our approach successfully enhances model robustness and class equilibrium.

To qualitatively validate the effectiveness of the proposed symmetrical modules, we visualized the decision-making process using Grad-CAM, as shown in Figure 10. In the baseline YOLOv8n (Row b), the model exhibits significant ‘attention drift.’ Specifically, in Viral Pneumonia cases, the activation maps heavily align with the rib cage structures rather than the lung parenchyma, confirming our hypothesis that standard convolutions struggle to decouple overlapping anatomical noise. Similarly, in Bacterial cases, the baseline is distracted by high-contrast regions like the clavicles and mediastinum. In contrast, the proposed YOLO-SMD (Row c) demonstrates superior ‘attention concentration. Thanks to the SAC2f module, the model successfully suppresses the structural noise of the ribs, directing focus solely to the pathological interstitial patterns in viral cases. Furthermore, the SDFM unit ensures that the focal consolidations in bacterial pneumonia are accurately targeted without background leakage. This visual evidence confirms that our architectural improvements fundamentally alter how the model perceives lesions, shifting from edge-based coincidences to true pathological feature extraction.

Figure 11 presents the detection results of various models on typical chest X-ray images of bacterial and viral pneumonia. Overall, all models successfully identify paired lesion boxes in the bilateral lung regions, demonstrating a certain degree of object recognition capability. However, significant differences are observed in terms of confidence and box localization accuracy. YOLOv8n and YOLOv10n typically predict bacterial and viral pneumonia with confidence scores ranging from 0.7 to 0.80, with some boxes being slightly oversized and covering a substantial portion of the normal lung fields. YOLOv11n and RT-DETR-r18 show cases where the confidence for boxes on one side is notably lower, highlighting issues with localization stability in certain samples. In contrast, the proposed method achieves higher confidence, typically around 0.85 for bacterial pneumonia samples, and maintains a confidence range of approximately 0.81–0.87 for viral pneumonia samples. Furthermore, the predicted boxes are more accurately aligned with the actual lesion areas, exhibiting clearer boundaries and reduced background redundancy. These visual detection improvements align with the quantitative gains in mAP and Recall, demonstrating that the proposed method provides more precise lesion annotations with greater confidence for both types of pneumonia.

Figure 12 presents the PR curves for YOLOv8n and our method on Bacterial Pneumonia, Viral Pneumonia, and the overall categories. Overall, both methods maintain high precision in the low recall region, with precision gradually decreasing as recall increases. However, the three curves of our method remain slightly higher than the baseline across most recall intervals, indicating that our approach achieves higher precision at the same recall levels. Numerically, the baseline model achieves AP values of 0.843 and 0.803 for Bacterial Pneumonia and Viral Pneumonia, respectively, with an overall mAP@0.5 of 0.823. In contrast, our method improves these values to 0.862 and 0.831, with an overall mAP@0.5 of 0.846, showing improvements across all metrics. This demonstrates that, while maintaining recall, our method further reduces the false positive rate for both types of pneumonia, achieving a superior precision-recall trade-off.

5.3. Ablation Experiment

To quantitatively evaluate the contribution of each improvement module to the overall detection performance, this study uses YOLOv8n as the baseline model and incorporates components such as C2f-Star-CAA, SDFM, and DySample. Ablation experiments are conducted under consistent experimental conditions. The evaluation metrics—Recall, mAP@0.5, and mAP@0.5:0.95—are used to analyze the impact of each module both individually and in combination on the model’s detection performance. The detailed results for each configuration are presented in

Table 9. Results of Ablation Experiment.

Number	Experiments	Recall	mAP@0.5	mAP@0.5:0.95
1	YOLOv8n	85.0	82.3	50.4
2	YOLOv8n + C2f-Star-CAA	82.4	84.0	52.0
3	YOLOv8n + SDFM	84.2	84.1	52.1
4	YOLOv8n + DySample	83.3	83.5	51.4
5	YOLOv8n + SDFM + C2f-Star-CAA + DySample	86.1	84.3	52.6

.

As shown in Table 9, compared to the baseline YOLOv8n, the introduction of C2f-Star-CAA (Experiment 2) and SDFM (Experiment 3) individually leads to significant improvements in precision-related metrics. Specifically, mAP@0.5 increases to 84.0% and 84.1%, while mAP@0.5:0.95 rises to 52.0% and 52.1%, indicating that both modules effectively enhance feature representation and localization accuracy. However, this improvement is accompanied by a slight decrease in Recall or performance comparable to the baseline. Specifically, as shown in Experiment 4, the introduction of the DySample module alone yields an mAP@0.5 of 83.5%, which is a 1.2% improvement over the baseline. This result validates that the content-aware upsampling mechanism effectively reconstructs fine boundary details, thereby enhancing localization precision. When SDFM, C2f-Star-CAA, and DySample are combined (Experiment 5), Recall increases to 86.1%, while mAP@0.5 and mAP@0.5:0.95 improve to 84.3% and 52.6%, respectively, achieving optimal results across all three metrics. This demonstrates a strong synergistic effect among the modules, which not only enhances detection precision but also boosts Recall, leading to a significant overall improvement in detection performance. Crucially, these empirical results support our theoretical justification: the removal of the spatial reconstructor (DySample) leads to boundary loss, while the absence of the cross-dimensional modulator (SAC2f) or the semantic gate (SDFM) degrades feature discrimination. This validates that the chosen modules are not merely additive enhancements but are essential components for restoring the structural symmetry required for pediatric pneumonia detection.

6. Discussion

6.1. Pathophysiological Interpretation of Model Performance

The quantitative superiority of YOLO-SMD, as evidenced in Section 5, is not merely a result of increased parameter complexity but stems from the precise alignment between our symmetrical design philosophy and the anatomical constraints of pediatric physiology. Unlike general object detection, pediatric pneumonia screening requires overcoming specific biological barriers that standard models often fail to address.

6.1.1. DySample: Counteracting Alveolar Immaturity

Pediatric lungs are anatomically distinct from adults; the alveoli are smaller and less developed, and the collateral ventilation channels (pores of Kohn) are immature. Consequently, pneumonia in children—especially viral cases—often presents as “ground-glass” opacities with extremely diffuse gradients rather than the solid, well-defined consolidations seen in adults. Standard upsampling methods (Nearest Neighbor/Bilinear) essentially apply a low-pass filter, smoothing these faint gradients and causing “feature washout” of early-stage lesions. The deeper insight here is that DySample acts as a “structure-preserving reconstructor.” By generating content-aware offsets, it essentially “pulls” the sampling grid towards the subtle intensity shifts in the lesion boundary. This allows the model to retain high-frequency edge information even when the anatomical boundary is biologically hazy, directly resolving the “Boundary Ambiguity” inherent in immature lung tissue.

6.1.2. SAC2f: Decoupling Lesions from Thoracic Structural Noise

A unique challenge in pediatric X-rays is the high cardiothoracic ratio and the density of the rib cage relative to the lung volume. This creates significant “structural noise” where ribs and the heart often overlap with lung fields, mimicking or obscuring lesions (e.g., retrocardiac pneumonia). Standard CNN blocks often struggle to differentiate the “repetitive edge patterns” of ribs from the “irregular texture patterns” of pneumonia. The SAC2f module, through its Star-CAA mechanism, introduces cross-dimensional interaction. This allows the network to learn not just where a feature is (spatial) but what it represents (channel) simultaneously. By enforcing this symmetry, the model effectively decouples the pathological signal (lesion) from the anatomical background (ribs/heart), explaining the significant reduction in false positives in the hilar and basal regions.

6.1.3. SDFM: Adapting to the Viral-Bacterial Spectrum

Pediatric pneumonia represents a broad pathological spectrum: Viral Pneumonia is typically interstitial (texture-dominant, scattered), while Bacterial Pneumonia is alveolar (shape-dominant, consolidated). A static feature fusion strategy (like PANet) forces the network to compromise between these two conflicting modalities. The SDFM’s gated fusion functions as a “dynamic clinical triage” mechanism at the feature level. For viral cases, the gates suppress deep semantic shape features (which are irrelevant for scattered textures) and amplify shallow textural cues. For bacterial cases, the reverse occurs. This explains why our model achieves a balanced high performance across both categories (Table 4), as the architecture structurally mirrors the radiologist’s cognitive process of shifting attention between “texture analysis” and “shape recognition” based on the infection type.

6.2. Analysis of Performance Superiority Based on Symmetrical Design

The experimental results presented in Section 5 demonstrate that YOLO-SMD achieves a superior balance between Recall (86.1%) and mAP (84.3%) compared to state-of-the-art baselines like YOLOv12n and RT-DETR. This performance gain is not merely numerical but stems from the architectural symmetry designed to address the specific pathologies of pediatric pneumonia. Standard detectors (e.g., YOLOv10/v11) often fail on pediatric X-rays because they treat feature extraction uniformly. However, our Gap Analysis identified that pediatric lesions exhibit “boundary ambiguity” and “extreme scale variation.”

• Resolution of Boundary Ambiguity: The superior performance in mAP@0.5:0.95 (52.6%) confirms that the DySample module effectively counters the aliasing effects of standard Nearest Neighbor upsampling. By dynamically modulating sampling points based on content, the model “sharpened” the blurred edges typical of immature lung tissue, which standard YOLO models failed to delineate precisely.
• Resolution of Semantic Gaps: The significant boost in Bacterial Pneumonia AP (86.3%) validates the SDFM design. By using a gated fusion mechanism rather than simple concatenation, the model successfully synthesized deep semantic features (crucial for identifying large bacterial consolidations) with shallow textural cues, enforcing a symmetrical representation of multi-scale features.

Evidence of Symmetry-Based Deviation Learning: The hypothesis regarding the link between architectural symmetry and anatomical detection is empirically validated by our observations:

• Suppression of Symmetric Structural Noise: As observed in the Grad-CAM visualizations (Figure 10, row b vs. c), standard asymmetric detectors often exhibit “attention drift” towards the rib cage. Since ribs are bilaterally symmetric and edge-rich, an asymmetric network falsely amplifies them. By enforcing dimensional symmetry via SAC2f, YOLO-SMD treats spatial location and texture content equally, allowing the model to “filter out” the symmetric rib patterns and isolate the asymmetric lesions.
• Sensitivity to Symmetry-Breaking: The ablation study (Table 9) confirms that when the symmetrical interaction (SAC2f) or gated balance (SDFM) is removed, the model’s Recall drops significantly. This quantitative decline serves as empirical proof that Architectural Symmetry is a prerequisite for robustly detecting the “symmetry-breaking” events of pneumonia in a noisy anatomical environment.

6.3. Implications for Deployment in Primary Healthcare

The ultimate goal of this study is to assist diagnosis in primary healthcare settings, where expert radiologists are scarce and hardware resources are limited.

• Hardware Feasibility: The proposed YOLO-SMD maintains a lightweight architecture. The analysis in Section 6.2 confirms that the additional modules (SAC2f, DySample) do not create a bottleneck for deployment. This implies that the model can be deployed on mid-range edge devices or standard medical workstations without requiring expensive high-performance computing clusters, which is essential for low-resource regions.
• Clinical Validity (Recall vs. Precision): In primary screening, the cost of a “missed diagnosis” (False Negative) is significantly higher than a “misdiagnosis” (False Positive). Our method achieves the highest Recall of 92.1% for Viral Pneumonia (Table 4), outperforming YOLOv12n by 2.3%. This high sensitivity acts as a robust “safety net” for junior doctors, ensuring that subtle, occult viral lesions are flagged for further review.
• Symmetry in Clinical Workflow: The model’s balanced performance across both Bacterial (shape-dominant) and Viral (texture-dominant) pneumonia ensures it can serve as a versatile tool in the initial triage of pediatric patients, supporting the clinical decision-making process where mixed infections are common.

Integration with PACS and Clinical Workflow

To translate the theoretical speed of YOLO-SMD into practical clinical value, the model must be seamlessly integrated into the hospital’s existing ecosystem, specifically the Picture Archiving and Communication System (PACS). We propose a standard-compliant deployment workflow designed to minimize disruption to the radiologist’s routine:

(1)

DICOM Interoperability: The deployment framework is built upon the DICOM (Digital Imaging and Communications in Medicine) standard. The YOLO-SMD inference engine functions as a distinct DICOM node.

• Input: Chest X-rays are automatically routed from the acquisition modality (CR/DR scanners) to the AI server via the C-STORE protocol.
• Processing: The system parses the pixel data, normalizes the bit-depth (e.g., 12-bit to 8-bit), and performs inference using the proposed YOLO-SMD model.
• Output: Detection results are not just saved as raw images but are encapsulated as DICOM Secondary Capture (SC) objects or DICOM Structured Reports (SR). The bounding boxes are burned into a separate layer or stored as coordinate metadata, allowing radiologists to toggle the AI overlay on or off at their viewing workstation.

(2)

End-to-End Latency Analysis: While our model achieves an inference latency of 1.9 ms (approx. 526 FPS) on a GPU, the “System Turnaround Time” perceived by the clinician includes data transmission and DICOM I/O overhead.

• In a typical local area network (LAN) environment within a hospital (bandwidth > 1 Gbps), the transmission of a high-resolution X-ray (approx. 10–15 MB) takes roughly 100–200 ms.
• Combined with preprocessing and DICOM re-packaging, the total time from image acquisition to result availability is estimated to be under 1 s.
• This near-instantaneous processing ensures that the AI analysis is already available by the time the radiologist opens the study, creating a zero-wait workflow.

(3)

Privacy and On-Premise Deployment: To adhere to data privacy regulations (e.g., HIPAA or GDPR), the lightweight nature of YOLO-SMD (Low GFLOPS) allows for On-Premise Deployment. The model can run on edge servers directly within the hospital’s firewall, eliminating the need to upload sensitive pediatric patient data to external cloud services, thus ensuring data security while maintaining high availability.

6.4. Limitations and Failure Analysis

While YOLO-SMD demonstrates promising results, it is imperative to acknowledge the limitations inherent in the algorithm, data, and evaluation environment to guide future clinical translation.

6.4.1. Dataset Bias and Annotation Uncertainty

The experimental data used in this study is sourced from the publicly available Chest X-Ray Pneumonia Dataset [26], which was initially compiled and released by Kermany et al. Although widely used, this reliance introduces specific biases:

• Demographic Bias: The dataset represents a specific regional population. Disease patterns and anatomical features may vary across different ethnic groups or geographic regions, potentially limiting the model’s direct transferability to non-Asian populations.
• Annotation Uncertainty: The ground truth labels used for training are based on radiological consensus (“silver standard”) rather than microbiological or pathological confirmation (“gold standard”). Given the inherent subjectivity in interpreting pediatric chest X-rays—where viral pneumonia often presents as subtle, non-specific interstitial changes—there exists an irreducible “aleatoric uncertainty” in the labels. Consequently, the model may inherit these human biases, potentially overfitting to the specific stylistic interpretations of the annotating radiologists. While strict filtering ensured high data quality, the limited sample size (n = 2611) raises concerns regarding generalization. To assess whether our architectural improvements were overfitting to this specific medical dataset, we conducted cross-domain validation on the VisDrone dataset (Table 8). The consistent performance gains observed in VisDrone demonstrate that the architectural generalization of YOLO-SMD is robust—effective at detecting small, dense objects regardless of the domain. However, medical generalization remains a challenge. As noted in the limitations, the current model has not yet been validated on external cohorts from different hospitals (Cross-Center Validation), which is necessary to rule out device-specific overfitting. Future work will prioritize multi-center data collection.

6.4.2. Algorithmic Failure Modes

Despite the inclusion of DySample and SAC2f to handle blurred boundaries, likely failure modes persist in complex clinical scenarios:

• Comorbidities and Artifacts: The current model was trained primarily on clear pneumonia cases. It may struggle with “noisy” radiographs common in ICUs, such as those containing electrocardiogram leads, intubation tubes, or foreign bodies. Additionally, comorbidities like congenital heart disease (cardiomegaly) or pleural effusion may obscure lung fields, leading to false negatives or mislocalization.
• Motion Artifacts: Pediatric patients are often uncooperative, leading to motion blur. While DySample improves feature alignment, extreme blurring caused by rapid breathing or movement may still result in missed detections for small, low-contrast viral lesions.

Anatomical Mimicry and False Positives: Beyond external artifacts, the model’s remaining errors are predominantly driven by intrinsic anatomical structures acting as “mimics”:

• Rib Crossings (Pseudo-Nodules): The most frequent source of false positives for Viral Pneumonia occurs at the intersection of anterior and posterior ribs. These osseous overlaps create focal areas of increased density that morphologically resemble small pulmonary nodules. Although the SAC2f module suppresses much of this structural noise, the model still occasionally flags high-contrast rib crossings as lesions, particularly in the apical regions.
• Hilar and Cardiac Shadows: For Bacterial Pneumonia, false positives tend to cluster around the lung hilum and the cardiac silhouette. The confluence of major pulmonary vessels and bronchi at the hilum can mimic the appearance of central consolidation. Similarly, the right heart border, if prominent or slightly rotated, can be misinterpreted by the model as a lobar infiltrate. Future integration of explicit anatomical segmentation (e.g., rib suppression bone-suppression algorithms) could further mitigate these errors.

6.4.3. Evaluation Environment vs. Deployment Reality

Our evaluation was conducted on a high-performance NVIDIA RTX 4090 GPU (Section 4.2). While the model is theoretically “lightweight” (low parameter count), the inference latency reported in this ideal environment may not perfectly reflect performance on resource-constrained hardware found in primary care clinics (e.g., older CPUs or embedded devices like Jetson Nano). The gap between “algorithm complexity” and “real-world throughput” necessitates future work on model quantization (FP16/INT8) to ensure true accessibility.

6.4.4. Scope of Clinical Application

A limitation of the current experimental design is the exclusion of “Normal/Healthy” chest X-rays. The proposed YOLO-SMD is trained specifically to distinguish between bacterial and viral pathologies in patients with radiographically visible abnormalities. Consequently, the model’s behavior on completely healthy lungs has not been quantified in this study. Therefore, in clinical practice, this tool is intended to be used as a “Second-Stage Classifier” for patients with confirmed pulmonary opacity, rather than a “First-Stage Screener” for the general healthy population. Using the model on healthy subjects could potentially lead to false-positive lesion detections (hallucinations) due to the lack of negative samples in the training distribution. Future work will involve integrating a “Normal” class to extend the model’s capability to primary screening.

In terms of qualitative evaluation, the current study primarily employed visualization examples that are representative and clear in pathology to illustrate the model’s core mechanisms—such as attention concentration and boundary reconstruction—under relatively controlled imaging conditions. While these examples effectively demonstrate the model’s ability to handle typical pneumonia manifestations, they do not encompass borderline or diagnostically ambiguous cases that are frequently encountered in clinical practice. The inclusion of such challenging cases would undoubtedly enhance the interpretability and clinical relevance of the model. However, given the scope of this work—which focuses on architectural innovation and quantitative benchmarking—we have limited the qualitative analysis to a concise set of illustrative samples. Future studies will systematically expand the qualitative evaluation to include a wider spectrum of cases, particularly those with diagnostic uncertainty, to better assess the model’s robustness and decision boundaries in real-world clinical scenarios.

6.4.5. Quantitative Analysis of Computational Trade-Off

A critical concern for medical AI deployment is the trade-off between diagnostic accuracy and computational cost. Although YOLO-SMD remains lightweight, the introduction of DySample and attention modules incurs a marginal increase in computational load. To justify this, we conducted a quantitative “Marginal Utility” analysis comparing YOLO-SMD against the baseline YOLOv8n.

(1)

Quantitative Cost–Benefit Analysis:

According to the experimental data, the computational cost (GFLOPS) increased marginally from 8.7 to 8.9 (a relative increase of roughly 2.3%), and the inference latency rose by 0.6 ms per image. In exchange, the model delivers a 2.4% relative improvement in mAP@0.5 (from 82.3% to 84.3%) and a 2.4% relative improvement in Recall (from 83.7% to 86.1% compared to YOLOv12n).

We define the Efficiency Ratio (ER) as the quotient of relative performance gain to relative computational cost. Our model achieves an ER of approximately 1.04 for mAP, indicating a favorable “super-linear” return on investment: every 1% increase in computational budget yields more than 1% gain in detection accuracy.

(2)

Clinical Real-Time Validity:

While the latency shows a statistical increase, it must be contextualized within clinical requirements. The inference speed of our model is approximately 1.9 ms, which corresponds to 526 Frames Per Second (FPS). Since standard medical imaging displays and real-time fluoroscopy typically operate at 30–60 FPS, the drop from the baseline’s 769 FPS to 526 FPS is imperceptible to the human eye and effectively zero-cost in a practical user experience.

6.5. Comparison with Clinical Diagnostic Thresholds (PPV and NPV)

To further contextualize the model’s performance from a clinical decision-making perspective, we consider the Positive Predictive Value (PPV) and Negative Predictive Value (NPV). In our binary detection framework (lesion present vs. absent within an image), PPV is equivalent to Precision, while NPV depends on both Recall (Sensitivity) and the prevalence of radiographic pneumonia in the target population. Based on our test set results:

• At the high-recall operating point (prioritizing detection of viral pneumonia, Recall = 92.1%, Precision = 78.4%), the model achieves a high NPV. Assuming a moderate prevalence, this translates to a very low probability of pneumonia when the model predicts “negative,” making it a reliable tool for “ruling out” disease in a screening context—a critical feature for primary care where missed diagnoses are a major concern.
• Conversely, at a high-precision operating point (e.g., for confirming bacterial consolidation, Precision > 86%, Recall = 80.1%), the model delivers a high PPV, indicating a high probability of disease when the prediction is positive. This supports its use as a triage tool to flag cases with high confidence for urgent review.

The choice between these operating points can be adjusted via the detection confidence threshold based on clinical priorities. The proposed model’s balanced performance across both recall- and precision-centric metrics (as shown in the PR curves, Figure 10) provides flexibility for such adaptation. It is important to note that precise PPV/NPV estimates require knowledge of the true disease prevalence in the deployment population. Our current dataset, focused on confirmed pneumonia cases, does not include “normal” X-rays, thus limiting our ability to calculate population-specific NPV/PPV. Future validation on a fully representative clinical cohort, including healthy subjects, is necessary to establish definitive clinical utility thresholds. Nevertheless, the achieved high recall directly supports a high NPV, which is paramount for a safety-net screening tool in resource-limited settings.

6.6. Addressing Domain Shift Across Pediatric Datasets

A key challenge in applying our model in diverse clinical settings is the domain shift across different pediatric datasets, such as the PERCH and Guangzhou datasets. These datasets may contain variations in imaging protocols, disease manifestations, and patient demographics, all of which can impact the model’s generalization ability. For example, the differences between datasets like PERCH and the Chest X-ray Pneumonia Dataset (Kermany et al. [26]), which was used for training in this study, can lead to performance inconsistencies when transferring the model to new data sources.

To mitigate this issue, we plan to incorporate domain adaptation techniques in future iterations of the YOLO-SMD framework. By fine-tuning the model on different pediatric datasets, we aim to improve its robustness and adaptability. Domain adaptation methods, such as adversarial training and feature normalization, could be explored to align feature distributions across datasets, thus minimizing performance degradation when moving between datasets. This would ensure that our model is better suited to handle the variations inherent in real-world, multi-domain clinical environments.

We believe that addressing domain shift is essential for enhancing the model’s practical applicability, particularly in primary healthcare settings, where resource constraints often limit the availability of high-quality annotated data. Ensuring the model’s robustness across different datasets will help establish its clinical validity and extend its usage to a broader range of pediatric pneumonia cases worldwide.

6.7. Clinical Implications of the Recall-Precision Trade-Off

While our proposed YOLO-SMD prioritizes Recall to function as a “safety net” in primary screening, we must address the clinical implications of the consequent increase in False Positives. In a clinical setting, an elevated False Positive Rate (FPR) can theoretically lead to “alarm fatigue” for radiologists, unnecessary follow-up radiation (e.g., CT scans), or parental anxiety.

However, we argue that this trade-off is clinically justified for pediatric pneumonia screening due to Risk Asymmetry. The “cost” of a missed diagnosis (False Negative)—which may result in a child with occult pneumonia deteriorating at home—is significantly higher than the “cost” of a False Positive, which typically results in a secondary review by a senior physician who can rule out the disease.

Furthermore, the model’s high recall is particularly pronounced for Viral Pneumonia (92.1%). From an Antibiotic Stewardship perspective, a false positive “Viral” prediction is less detrimental than a false positive “Bacterial” prediction. A viral flag typically prompts supportive care and observation rather than immediate antibiotic prescription. Therefore, the model avoids encouraging antibiotic abuse even when it leans towards over-sensitivity. Ultimately, the YOLO-SMD is designed for a Human-in-the-loop workflow: the AI maximizes sensitivity to ensure no lesion is overlooked, while the supervising clinician acts as the specificity filter to reject false positives.

7. Conclusions

This study addresses the critical challenges of boundary ambiguity and semantic gaps in pediatric pneumonia detection by proposing YOLO-SMD. Through the integration of SAC2f, SDFM, and DySample, the framework enforces a symmetrical design philosophy that balances deep semantic context with shallow textural details.

Experimental results on a curated dataset of 2611 images demonstrate that YOLO-SMD offers a clinically valuable optimization over state-of-the-art baselines. While the improvement in general detection accuracy (mAP@0.5) is modest (+0.3% compared to YOLOv12n), the model achieves a statistically significant gain in Recall (+2.4%), reaching 86.1%. Specifically, the model shows distinct strength in identifying viral pneumonia (Recall 92.1%). This indicates that rather than establishing a broad new benchmark, YOLO-SMD functions specifically as a high-sensitivity “safety net,” effectively reducing false negatives in primary screening scenarios.

However, the capabilities of this study must be interpreted within its limitations. First, the reliance on single-center data introduces potential demographic bias, and the lack of external validation across different scanner manufacturers limits our understanding of the model’s cross-domain robustness. Second, annotation uncertainty remains a challenge; as the model is trained on radiological labels rather than pathological biopsy results, it is susceptible to inter-observer variability. Third, the model is currently untested against real-world noise, such as motion artifacts or the presence of life-support devices, which are frequent in pediatric ICUs.

To address the limitation of data diversity inherent in single-center studies, future work will explicitly prioritize the expansion of the dataset by incorporating chest X-rays from varied sources, including different hospitals and scanner types. This multi-center external validation will be crucial to quantify and enhance the model’s robustness and generalizability before clinical deployment. Moving beyond dataset expansion, we will further investigate advanced domain adaptation techniques. This includes feature alignment methods (e.g., adversarial training or moment matching) to minimize distribution shifts between data from different hospitals or scanner types, and self-supervised contrastive learning strategies to learn more generalizable representations from unlabeled multi-center data. Additionally, techniques like Test-Time Adaptation (TTA) or Adaptive Batch Normalization (AdaBN) could be employed to dynamically adjust the model to new clinical environments during inference. These approaches collectively aim to systematically improve the model’s generalizability and reliability for deployment across varied healthcare settings. Ultimately, YOLO-SMD is positioned not as a standalone diagnostician, but as a robust, sensitivity-oriented auxiliary tool designed for integration into tiered medical systems. For deployment in primary healthcare settings, we recommend a two-stage workflow: First, a high-sensitivity screening model (like YOLO-SMD) flags potential abnormalities with high recall, particularly for subtle viral cases. Second, flagged cases or low-confidence predictions are escalated for review by healthcare workers or tele-radiologists. This approach optimally balances the need for early detection with the constraint of limited specialist availability. Furthermore, to enhance the model’s clinical applicability, future efforts will include systematic robustness evaluation against common image perturbations—such as noise, blur, and brightness variations—frequently encountered in real-world imaging conditions. This will provide a more comprehensive assessment of the model’s stability under diverse and challenging clinical scenarios. To facilitate this, future work will prioritize model optimization for mobile and edge deployment, including quantization (FP16/INT8) and pruning to further reduce the computational footprint and enable real-time inference on low-power devices commonly found in primary healthcare settings. Simultaneously, we will explore federated learning techniques to adapt the system to diverse local data distributions while maintaining patient privacy. This explicit focus on edge compatibility ensures that YOLO-SMD can transition from a research prototype to a practical, point-of-care screening tool.