Multi-scale deformable attention fusion network with global context modeling for chest X-ray lesion segmentation

Abstract

Chest X-ray segmentation is a critical technique for the screening and diagnosis of common thoracic diseases such as pneumonia and COVID-19, providing essential support for clinical diagnosis and quantitative analysis. However, the inherent challenges of chest X-ray images, including blurry lesion boundaries and irregular contours, pose significant difficulties for accurate segmentation. To address this problem, we propose a Multi-Scale Deformable Attention Fusion Network (MSDAFNet) for lesion segmentation in chest X-rays. MSDAFNet adaptively perceives lesion regions of varying shapes and sizes, thereby enhancing the model’s robustness to diverse pathological tissues. Specifically, we design a multi-scale deformable convolution attention module that employs channel grouping and dilated convolutions with different dilation rates to capture multi-scale feature maps without increasing parameters. Learnable offsets are introduced, allowing the convolution kernels to adaptively adjust sampling positions. Furthermore, considering the clustered distribution characteristics of small target lesion regions, a memory module is proposed before the decoder.This module establishes semantic correlations between image regions by introducing learnable key-value pairs and an attention mechanism, thereby supplementing convolutional features with the capability of modeling long-range dependencies.Extensive experiments conducted on two large-scale medical datasets, QaTa-COV19-v1 and QaTa-COV19-v2, demonstrate that MSDAFNet achieves higher accuracy and robustness in anatomical structure segmentation than existing methods, delivering competitive performance.

Introduction

In clinical practice, Chest X-ray (CXR) imaging, with its core advantages of convenient operation, low cost, and low radiation dose, has become the most widely used imaging tool for the screening and diagnosis of common thoracic diseases such as pneumonia, tuberculosis, and COVID-19 [1–3]. Accurate segmentation of key anatomical structures, including the lung parenchyma and lesion regions, in chest radiographs is a crucial step for the automated quantitative assessment of diseases. It is also a fundamental prerequisite for building reliable computer-aided diagnosis systems and supporting disease staging and treatment efficacy monitoring [4, 5]. The accuracy of this segmentation directly determines the reliability of subsequent lesion quantification, pathological analysis, and treatment response evaluation. However, the precise identification and delineation of lesion areas in chest radiographs currently rely heavily on the subjective judgment of radiologists. This approach is not only time-consuming and labor-intensive but also suffers from inconsistent interpretation, making it difficult to meet the demand for rapid diagnosis in scenarios involving large-scale screening or public health emergencies [6]. Concurrently, with the exponential growth of medical imaging data, the demand for efficient processing and precise analysis of chest radiographs is increasingly urgent. The challenge of rapidly locating lesion areas and extracting quantitative features from a large volume of CXR images has become a key obstacle to improving the diagnostic efficiency of thoracic diseases [7]. Therefore, the development of automatic anatomical segmentation technology for chest X-ray imaging, aimed at accurately extracting anatomical structures and locating lesions, is of great significance for enhancing both diagnostic efficiency and accuracy.

Despite notable progress in chest radiograph segmentation technology in recent years, several challenges persist in practical applications:

1. Large Scale Variation and Blurry Boundaries: Lesion sizes in CXR vary drastically, ranging from tiny nodules to extensive diffuse opacities. Furthermore, the intensity contrast between lesion regions and surrounding normal tissue is often low, resulting in inherently indistinct boundaries (as shown in Fig. 1(a)).

2. Irregular Contours: Chest X-ray images are two-dimensional projections where anatomical structures such as lung parenchyma, mediastinum, and blood vessels overlap. Lesion contours also frequently exhibit irregular morphological shapes (as shown in Fig. 1(b)).

Fig. 1.

Visualization of segmentation results. The red areas indicate over-segmentation, and the blue areas indicate under-segmentation. In (a), the red box highlights mis-segmentation and missed segmentation caused by blurry lesion boundaries, while the green box shows regions with clear boundaries

Therefore, fully leveraging the global semantic information in CXR images to model anatomical structural relationships is highly valuable for improving performance in real-world applications. Traditional methods for CXR segmentation have predominantly relied on classical image processing techniques such as thresholding, region growing, and active contour models [8–10]. While these methods have alleviated the manual burden to some extent, they are generally sensitive to noise and exhibit limited generalizability in complex clinical scenarios characterized by blurry tissue boundaries and overlapping anatomical structures, struggling to cope with the high complexity of contemporary clinical images [11]. In contrast, deep learning-based approaches have demonstrated superior performance in complex segmentation scenarios. Encoder-decoder architectures, represented by the U-shaped structure and its attention-enhanced or nested variants such as SwinUNet [12], AttentionUNet [13], and Unet++ [14], have achieved remarkable segmentation accuracy in medical imaging tasks due to their efficient cross-level feature fusion mechanisms [15–18]. Despite the effective progress made by existing deep learning methods in chest radiograph segmentation, achieving high segmentation accuracy in complex clinical scenarios remains challenging. These challenges primarily stem from the following aspects:

• Limited Geometric Structure Modeling of Standard Convolutions: Constrained by the fixed sampling grid of standard convolutions, their regular sampling patterns struggle to adapt flexibly to the highly irregular and complex geometric morphologies of lesion contours in chest X-rays. This inherent limitation hinders the precise capture of fine anatomical details and complex boundary information in CXR images.

• Insufficient Adaptability in Cross-level Feature Fusion: Simple feature concatenation operations are inadequate for dynamically distinguishing and weighting features from different semantic levels. This prevents the model from effectively selecting beneficial detailed information for segmentation tasks and suppressing interference from background noise and irrelevant anatomical structures, especially when dealing with blurry tissue boundaries.

• Limited Effectiveness in Global Context Modeling: In practical clinical scenarios involving image noise and uneven contrast, existing methods often fail to establish accurate global relationships. This difficulty in reliably distinguishing between lesions and anatomically similar tissues leads to decreased model generalizability.

Consequently, it is crucial to accurately capture multi-scale lesion features and subtle boundary details while effectively utilizing global contextual associations to suppress background interference and adapt to individual anatomical variations [19, 20]. Based on the above analysis, this study proposes a lesion-oriented Multi-Scale Deformable Attention Fusion Network (MSDAFNet) for chest X-ray lesion segmentation. MSDAFNet is designed as a local-to-global representation framework that integrates deformable multi-scale feature aggregation, boundary-aware gated fusion, and compact memory-guided context modeling within a unified encoder-decoder architecture. The main methodological contribution of this framework lies in its lesion-oriented feature coordination strategy, which aims to simultaneously enhance lesion-scale adaptability, preserve weak boundary cues, and improve semantic discrimination between lesion regions and anatomically similar background structures. Specifically, the Multi-Scale Deformable Convolutional Attention module (MSDA) is embedded in the high-semantic decoding stage to combine multi-branch dilated convolution with deformable spatial sampling. By adaptively adjusting the receptive field according to local image structures, MSDA improves the representation of lesions with variable sizes, irregular contours, and blurry boundaries. This design enables the network to capture both fine-grained boundary details and broader contextual information, ranging from small localized opacities to large diffuse lesion regions. Subsequently, a boundary-aware gated fusion strategy is introduced into the skip connections between the encoder and decoder. Instead of directly concatenating or adding cross-level features, this strategy adaptively regulates the interaction between shallow spatial details and high-level semantic information. As a result, lesion-relevant boundary cues can be selectively enhanced, while interference from ribs, vessels, and low-contrast lung parenchyma is suppressed.Finally, a compact Global Context-aware Memory Module is incorporated at the bottleneck layer to strengthen long-range semantic consistency. By using learnable key-value memory prototypes, the module establishes global contextual associations across image regions and provides complementary semantic guidance for distinguishing true lesion areas from anatomically similar background tissues. Through the coordinated design of these three components, MSDAFNet forms a lesion-oriented segmentation framework that addresses the major challenges of chest X-ray lesion segmentation, including scale variation, boundary ambiguity, irregular morphology, and insufficient global contextual consistency.In summary, the main contributions of this work are as follows:

1. A lesion-oriented chest X-ray lesion segmentation framework, MSDAFNet, is proposed. The main methodological contribution lies in a feature coordination strategy that integrates deformable multi-scale representation, boundary-aware cross-level fusion, and compact memory-guided context modeling within a unified encoder-decoder architecture. This design aims to address lesion-scale variation, blurry boundaries, irregular contours, and weak contextual consistency in chest X-ray images.

2. A Multi-Scale Deformable Convolutional Attention module is developed to enhance boundary-sensitive lesion representation. By coupling multi-branch dilated convolution with learnable deformable sampling, the module adaptively aligns contextual receptive fields with lesions of different sizes and irregular shapes, enabling the network to capture both fine-grained boundary details and broader contextual information.

3. A boundary-aware gated skip fusion strategy is incorporated to improve encoder-decoder feature interaction. By generating gating weights from aligned encoder and decoder features, this strategy adaptively calibrates shallow skip features under high-level semantic guidance. It aims to preserve lesion-relevant boundary cues while reducing interference from ribs, vessels, and low-contrast lung parenchyma during cross-level fusion.

4. A Global Context-aware Memory Module is incorporated at the bottleneck layer to strengthen long-range semantic consistency. By using learnable key-value memory prototypes, the module provides global contextual guidance for distinguishing lesion regions from anatomically similar background tissues.

Related work

Medical image segmentation

Medical image segmentation, as a key technology in the field of medical image analysis [21–23], involves assigning accurate category labels to each pixel in an image. The most commonly used X-ray images, due to their single grayscale channel nature, often exhibit low contrast between different tissues and chaotic structural edge information. This makes the segmentation of disease structures in chest X-ray images more challenging compared to traditional medical image segmentation tasks [24, 25]. From the perspective of the development of deep learning-based medical image segmentation technologies, current methods for chest X-ray disease structure segmentation primarily follow two technical routes: methods based on Convolutional Neural Networks (CNNs) and methods based on Transformers.

Traditional U-shaped network architectures

Convolutional Neural Networks, with their powerful capability for local feature extraction, high computational efficiency, and adaptability to small sample data, have become the mainstream technical route for early chest radiograph lesion segmentation. Their core advantage lies in capturing key local features such as texture and edges through convolution operations, while effectively preserving fine-grained anatomical details via skip connections. Zhou et al. proposed U-Net++ [14], which optimizes feature fusion through a nested dense connectivity structure, strengthening the association between shallow detail features and deep semantic features. Similarly, Oktay et al. introduced Attention U-Net [13], which uses attention gates in skip connections to highlight task-relevant regions and suppress irrelevant encoder responses. Recent boundary-aware studies have further shown that attention-based feature refinement is beneficial for improving boundary discrimination in medical image segmentation [23]. Subsequently, the nnU-Net network proposed by Isensee et al. [26] significantly improved the generalizability and performance stability across different medical image segmentation tasks by automatically optimizing preprocessing pipelines, network architectures, and training strategies. Furthermore, Huang et al. proposed a method based on multi-scale interaction and deformable upsampling [27] to specifically improve the segmentation accuracy of the model for small target regions.

Improved methods based on transformers

The Transformer architecture, leveraging the long-range dependency modeling capability of the self-attention mechanism, effectively compensates for the shortcomings of CNNs in capturing global contextual information. Since Chen et al. proposed TransUnet [28] in 2021, Transformer-based segmentation methods have gradually become a research hotspot in the field of medical image segmentation. In the same year, Hu et al. introduced Swin-Unet [12], which uses a Swin Transformer based on sliding windows as the encoder. This approach maintains the ability to model long-range dependencies while reducing computational complexity through local window attention. Concurrently, Hatamizadeh et al. proposed UNETR [29], which deeply integrates a Transformer encoder with a U-Net decoder. By facilitating multi-level feature interaction, it strengthens the association between local details and global semantics, demonstrating strong potential in medical image segmentation tasks.Subsequently, Zhang et al. proposed ReT-UNet [30], which integrates a recursive feature aggregation block with a Transformer-based multi-scale global feature fusion module, balancing the capture of local details and global context, thereby enhancing the accuracy and robustness of segmentation in small-sample medical images. Additionally, Yan et al. proposed MDSTransUNet [31], which integrates ShuffleNet and Transformer modules in the encoder along with a dual-decoder branch to simultaneously segment lung structures and infected regions. These hybrid CNN–Transformer methods demonstrate that combining convolutional local feature extraction with Transformer-based global dependency modeling is an effective strategy for improving medical image segmentation performance.Nevertheless, Transformer-based and hybrid CNN–Transformer methods still face several challenges in chest radiograph lesion segmentation. First, although self-attention is effective for global semantic interaction, it usually introduces additional computational cost, especially when high-resolution feature maps are involved. Second, Transformer-based models often require sufficient training samples to fully exploit their global modeling ability, whereas pixel-level annotated chest X-ray lesion datasets are still relatively limited. Third, chest X-ray lesions are frequently characterized by low contrast, blurry boundaries, irregular morphology, and interference from ribs, vessels, and lung parenchyma. Under such conditions, global dependency modeling alone may not be sufficient to preserve weak lesion boundaries and local geometric details.Therefore, chest X-ray lesion segmentation requires a balanced representation strategy that can jointly capture global contextual relationships, local boundary details, and morphology-adaptive spatial features.

Multi-scale deformable convolution

In standard convolution kernels, the receptive field is typically a fixed rectangular region, which limits their sensitivity to geometric morphological features in CXR images [32]. To address this limitation, many studies have designed various variants of convolution operations to improve the adaptability of segmentation networks to morphological variations of target structures. Dilated convolution [33], as a classic method, expands the receptive field of the convolution kernel by setting different dilation rates, effectively enhancing the model’s ability to aggregate contextual information from targets of varying sizes. ASPP-like multi-scale context modules further extend this idea by combining several parallel dilated convolution branches to capture contextual information at different receptive-field scales. However, these methods mainly rely on predefined dilation rates, and their sampling positions remain fixed once the dilation patterns are determined. Therefore, although they can enlarge the contextual perception range, they still struggle to flexibly adapt to the complex contours and weak boundary transitions of irregular lesions in chest X-ray images. In contrast, deformable convolution dynamically learns offsets for each position in the feature map, allowing the receptive field to adaptively adjust to local image structures. This improves the adaptability of convolution kernels to complex lesion morphology. Based on this concept, Yang et al. proposed the DCU-Net model [34], which combines deformable convolution with a cascaded U-Net architecture. By flexibly adjusting the sampling positions of the convolution kernel, it better adapts to structures with complex shapes and multi-scale features. Meanwhile, Qi et al. proposed Snake Convolution, which differs from deformable convolution by introducing continuity constraints to optimize the receptive field adaptation process, effectively addressing the issue of receptive field deviation from target regions in the segmentation of long, curved tubular structures.Although existing methods have made considerable progress, they still face challenges in capturing features at the boundaries of blurry tissue structures. For instance, lesion tissues in CXR images often exhibit inherent characteristics such as a significant scale span from tiny nodules to diffuse opacities, and irregular, blurry boundaries [35]. Current methods may be limited when scale-aware contextual perception and morphology-adaptive spatial modeling are required simultaneously. Motivated by this requirement, the multi-scale deformable convolution module in this work integrates a multi-dilation-rate parallel convolution framework with deformable convolution, thereby combining the contextual perception capability of multi-scale dilated convolution with the adaptive sampling capability of deformable convolution. This design aims to improve the representation of CXR lesion tissues with scale diversity and morphological irregularity.

Method

Overall architecture

The MSDAFNet framework is based on the classic encoder-decoder architecture. Its design core lies in constructing a synergistic processing pipeline that encompasses global context enhancement, optimized feature fusion, and multi-scale detail refinement. The encoder module extracts multi-scale feature representations from the input chest X-ray image through progressive down-sampling. After encoding, the extracted features are progressively passed to the Memory Module. This module captures long-range dependencies, enabling effective modeling of global contextual features. The generated feature map is then directly used as the bottleneck layer input for the decoder. To promote cross-level information fusion and dynamically select and weight multi-scale features from the encoder, a gated attention mechanism is introduced in the skip connection stage. The decoder gradually restores spatial resolution through multi-stage up-sampling. A Multi-scale Deformable Convolution Attention Module is embedded at key stages. This module adaptively fits target morphology through deformable convolution and enhances local detail representation via a parallel multi-branch structure and attention mechanisms. Finally, the network outputs a pixel-level probability distribution map, achieving high-precision medical image segmentation. The overall network architecture is illustrated in Fig. 2.

Fig. 2.

Overall architecture of MSDAFNet

Encoder

A pre-trained mobileNetV3-large [36] backbone is employed as the encoder to leverage its efficient feature extraction capability. The input chest radiograph is transformed into a hierarchical multi-scale feature pyramid. Specifically, the encoder extracts features sequentially from five key stages, progressively reducing the spatial resolution to 1/2, 1/4, 1/8, 1/16, and 1/32 of the input. In this feature hierarchy, the shallow features retain higher spatial resolution, focusing on capturing detailed structures such as edges and textures. The deeper features exhibit a significantly higher level of semantic abstraction, making them more suitable for carrying global contextual information. This representation ranging from fine-grained to coarse-grained provides a structured, multi-level information foundation for subsequent decoding and fusion processes.

Memory module

In traditional neural networks, expanding the receptive field often relies on image down-sampling strategies [37]. By progressively reducing the feature map size through pooling operations or strided convolutions, each neuron can cover a larger region in the input image. However, this approach leads to the loss of spatial details, localization ambiguity, and an inability to effectively model long-range dependencies [38]. To address this, a global context-aware Memory Module is introduced at the bottleneck layer following the encoder. A global context dictionary is constructed to supplement high-level convolutional features with global semantic information. Specifically, high-level features are passed through a 1×1 convolution to generate query keys. These keys are then matched for similarity with memory keys, acting as a global context enhancer. This design aims to address the difficulty in modeling long-range dependencies caused by scattered anatomical structures and significant variations in lesion scales within chest X-ray images. The module builds a learnable global memory bank to store a set of global prototype features, enabling feature association and information aggregation across spatial positions. The specific structure of the proposed Memory Module is shown in Fig. 3.

Fig. 3.

Structure of the memory module

Feature Mapping and Dimensionality Transformation: The deep feature map output by the encoder is passed through a 1 × 1 convolutional layer for dimensionality reduction and feature transformation, generating the query features . Simultaneously, the global memory bank within the module consists of a set of key-value pair parameter matrices: the memory keys and the memory values . Here, N represents the capacity of memory entries, and C\prime denotes the feature dimension. The key matrix is used to compute similarity with the input queries, while the value matrix stores the corresponding global contextual feature prototypes. Together, they collaborate to establish long-range dependencies across image regions.

Attention Weight Calculation: The query (with its spatial dimensions flattened) is matrix-multiplied with the transposed memory keys . Attention weights are then calculated using the Softmax function:

1

where is the feature dimension scaling factor. The attention weights reflect the association strength between each position in the input feature map and each global prototype in the memory values .

Global Information Aggregation and Output: The computed attention weights are used to perform a weighted sum over the memory values , aggregating global contextual information for each query position. The resulting enhanced features are reshaped back to a two-dimensional spatial structure. Finally, the output is added to the original input features via a residual connection, producing the final output feature map :

2

For clarity, Algorithm 1 provides the pseudo-code for the Memory Module.

Decoder

The decoder adopts a U-shaped hierarchical structure symmetrical to the encoder, consisting of five consecutive decoding blocks. Through four progressive up-sampling operations, the spatial resolution of the feature maps is gradually restored, ultimately generating a segmentation prediction map with the same dimensions as the input image. At each decoding stage, the deep semantic features from the previous layer are first up-sampled via bilinear interpolation to match the spatial dimensions of the corresponding feature map transmitted from the encoder via a skip connection.Subsequently, the two feature streams are fed into a boundary-aware gated fusion module for feature calibration. In this module, the up-sampled decoder feature provides high-level semantic guidance, while the encoder skip feature provides spatial details and boundary cues. The gating map is generated from the aligned encoder-decoder features and is used to calibrate the skip features before concatenation. This design aims to reduce background interference while preserving lesion-related weak boundary information during cross-level fusion. The structure of the gated fusion module is illustrated below:

As shown in the Fig. 4, independent 1×1 convolutions and batch normalization operations are first applied to the decoder’s up-sampled features and the encoder’s skip features for dimension alignment. The aligned features are then combined by element-wise addition to generate a joint semantic-spatial representation. After non-linear transformation, a Sigmoid activation function is used to obtain the gating weight map , as specified by the following formula:

Fig. 4.

Structure of the gated attention mechanism

where σ denotes the Sigmoid function, and represents the convolutional weights. Different from direct feature concatenation, this gating process introduces decoder-guided semantic constraints before skip fusion, allowing the network to selectively retain lesion-relevant encoder details and suppress redundant responses from anatomical background structures.

Finally, the gating map performs element-wise calibration on the aligned encoder features, producing the calibrated skip feature:

where represents the aligned encoder features, represents the transformed encoder features from the attention pathway, and ⊙ denotes element-wise multiplication. When the value of G is close to 1, the transformed skip features guided by the joint encoder-decoder representation are emphasized. When the value of G is close to 0, the original aligned skip features are retained. In this way, the module performs residual-style calibration of encoder features rather than directly replacing shallow details with decoder semantics. This helps preserve weak lesion boundaries while reducing irrelevant background activations.

Finally, the calibrated features are concatenated with the up-sampled decoder features along the channel dimension, followed by a set of standard convolutional layers for local fusion and smoothing. By repeating this process level by level, the spatial resolution of the feature maps is progressively restored, and their semantic representation and detailed information are gradually enriched.

Multi-scale deformable convolutional attention module

To enhance the representation of CXR lesions with scale variation, irregular morphology, and ambiguous boundaries, a Multi-Scale Deformable Convolutional Attention module is embedded in the high-level decoding stage of the network. In chest X-ray lesion segmentation, lesion regions may appear as small localized opacities or large diffuse abnormalities, and their boundaries are often affected by low contrast, anatomical overlap, and background interference. Therefore, the decoder requires not only multi-scale contextual perception, but also the ability to adapt its sampling pattern to local lesion morphology.

To capture contextual information at different spatial ranges, MSDA first adopts parallel convolutional branches with multiple dilation rates. Unlike conventional ASPP-like modules, which mainly enlarge the receptive field through fixed atrous convolutions, the proposed MSDA module further considers the spatial alignment between sampling locations and lesion morphology. Although fixed dilation patterns can provide scale-diverse contextual representations, their sampling positions remain regular and input-independent once the dilation rates are determined. For CXR lesions with irregular contours, weak boundary transitions, or discontinuous opacity distributions, such fixed sampling patterns may introduce irrelevant background responses or fail to sufficiently cover lesion-related boundary cues.

To address this limitation, MSDA introduces a deformable sampling path after multi-scale feature aggregation. Learnable offsets are predicted from local features and used to adjust the sampling locations of convolutional kernels. In this way, the receptive field is not only expanded by the multi-branch structure, but also adaptively adjusted according to the spatial distribution of lesion-related responses. This structure-adaptive sampling mechanism is beneficial for capturing small lesions, irregular contours, and blurry boundaries, especially in regions affected by ribs, vessels, or low-contrast lung parenchyma.

After deformable feature extraction, spatial attention and channel attention are further introduced to recalibrate the aggregated features. The spatial attention branch emphasizes lesion-relevant positions and suppresses redundant background responses, while the channel attention branch adaptively reweights different feature channels after multi-scale fusion. Through this coordinated design, the MSDA module provides a boundary-sensitive multi-scale representation for CXR lesion segmentation and enhances the model’s ability to capture lesions with variable sizes, irregular shapes, and ambiguous boundaries. The specific architecture of the MSDA module is illustrated in Fig. 5.

Fig. 5.

Architecture of the multi-scale deformable convolutional attention module

The specific architecture of the MSDA module consists of three components as follows:

Multi-scale feature extraction

For a given feature map output by the encoder, multi-scale features are first extracted through a set of parallel convolutional paths. These paths simulate receptive fields of different sizes using different dilation rates r [33]:

3

where denotes the output of the i-th branch. Lesion sizes in chest X-ray images typically range from 2 to 20 mm [39]. To better cover micro-nodules and large-scale lung field structures, dilation rates r∈[1,2,3,4] are selected to balance the perception capability for fine textures (small dilation rates) and broad contextual information (large dilation rates). After concatenating the outputs of all branches along the channel dimension, the fused multi-scale features are obtained:

4

Dynamic spatial adaptation

Offset Prediction: The offset prediction network is composed of two convolutional layers. First, a 1× 1 convolution reduces the channel dimension of multi-scale features to C//8. Subsequently, a 3× 3 convolution outputs offsets Δ P with dimensions [2–4], corresponding to the 2D (x, y) offsets for each position in the 3× 3 sampling grid of four grouped convolution kernels:

5

6

Specifically, the offset Δ P assigns a pair of coordinate offsets to each spatial position (i,j) of the feature map. This adjustment shifts sampling points from regular grid positions (x,y) to , guiding sampling points to move toward the edges and core regions of target structures. As a result, convolutional output features are more focused on meaningful target regions, and interference from irrelevant background information is suppressed.

Deformable Convolution and Spatial Weight Calculation: Traditional convolution adopts a fixed grid-based sampling pattern, where sampling positions strictly follow preset rules and are independent of input feature content. This makes it difficult to adapt to irregular lesion contours and blurry boundary characteristics in chest X-ray images. Although convolution kernels with fixed dilation rates expand the receptive field through preset parameters, they are still limited by these parameters and cannot dynamically respond to structural differences in different regions, leading to poor adaptability to lesions with variable shapes. As shown in Fig. 5, compared with the two traditional convolution methods mentioned above, deformable convolution breaks the inherent constraints of fixed sampling grids by introducing predicted offsets Δ P. It can adaptively adjust the spatial sampling positions and distribution of convolution kernels based on local structural information of the input feature map, thus accurately focusing on irregular target regions [40].

Deformable convolution employs a grouping strategy and adjusts sampling positions based on offsets Δ P to achieve adaptive fitting of irregular lesion structures. After obtaining the offsets Δ P from the previous step, deformable convolution is applied to the feature map X to generate , with the specific formula as follows:

7

where K=3 represents the size of the convolution kernel, W(k,l,c) denotes the weight parameter of the convolution kernel in the c-th channel, and is the feature map output by deformable convolution. Each channel of encodes sampled local structural information.

Since the features output by deformable convolution remain multi-channel features, channel aggregation is required to extract global responses in the spatial dimension, enabling the conversion from multi-channel features to single-channel attention weights. Global average pooling is performed on across channels to integrate multi-channel target responses, generating a single-channel spatial feature map. Finally, normalization to the range [0,1] is achieved through Sigmoid activation, resulting in spatial attention weights :

8

where denotes the Sigmoid activation function, and represents the output of the c-th channel of deformable convolution on the input feature map X modulated by offsets Δ P.

Attention mechanism and feature fusion

After obtaining spatial attention weights , the module achieves multi-level information integration through systematic feature fusion. First, element-wise multiplication is performed between spatial attention weights and the original input features X to realize adaptive weighting in the spatial dimension:

9

Subsequently, spatially enhanced features are concatenated with multi-scale features along the channel dimension to fuse local details and global contextual information, yielding . In this process, a channel attention mechanism is introduced to recalibrate the concatenated features across the channel dimension for further optimization. Specifically, global average pooling (GAP) is first applied to the input feature map to generate a channel statistical descriptor vector :

10

where each element in vector z represents the global response intensity of the c-th feature channel, and denotes the value of feature map at the c-th channel and spatial position (i, j). The vector z is then fed into a fully connected network to generate channel attention weights :

11

The obtained channel attention weights are multiplied with the concatenated features to achieve channel-selective feature enhancement. Finally, a 3× 3 convolutional layer is used to perform the final transformation on the enhanced features, outputting the refined feature map Y:

12

This fusion strategy ensures that multi-scale deformable convolution can simultaneously utilize the precise localization capability of spatial attention mechanisms, the contextual perception capability of multi-scale features, and the feature selection capability of channel attention mechanisms, significantly improving the representation accuracy of the model for complex structures in chest X-ray images.

For clarity, Algorithm 2 presents the pseudocode of the multi-scale deformable convolutional attention module.

Output layer

In the output layer, the high-dimensional feature map output from the decoder stage is compressed along the channel dimension by a 1×1 convolutional layer. The number of feature channels is directly mapped to the number of target classes, enabling a linear transformation from the feature space to the classification space. Subsequently, a Sigmoid activation function is applied to perform non-linear normalization on the output value of each pixel, converting it into a probability value within the range of [0,1], which represents the confidence level of each pixel belonging to the target region. Finally, bilinear interpolation is used to upsample the probability map to the original input resolution, ensuring spatial alignment between the output segmentation map and the input image. Throughout the output process, computational efficiency is guaranteed, while the multi-scale feature information extracted by the front-end network is completely preserved. A binary segmentation mask corresponding to each pixel of the input image is generated as the final output.

Loss function

In chest X-ray (CXR) lesion segmentation tasks, lesion regions typically account for only a small portion of the entire image. This severe class imbalance impairs the performance of traditional pixel-level loss functions. To address this issue, a weighted combination of Binary Cross Entropy (BCE) loss and Dice loss is adopted as the total training loss function for the network:

13

where N denotes the total number of pixels, represents the ground-truth label of the i-th pixel, and denotes the probability predicted by the model that the i-th pixel belongs to the lesion region.

14

where ε is a smoothing factor, which is used to prevent the denominator from being zero and enhance the stability of numerical calculations.

15

Experiment

Datasets

In this study, public chest X-ray image datasets jointly provided by Qatar University and Tampere University [41, 42] (QaTa-COV19-v1, QaTa-COV19-v2) were selected as the benchmark for experimental validation of all methods. The QaTa-COV19-v1 dataset contains 2951 posteroanterior (PA) chest X-ray images of laboratory-confirmed COVID-19 positive patients and their corresponding segmentation masks. As an extended version, the QaTa-COV19-v2 dataset includes 9258 chest X-ray images of COVID-19 positive patients and their corresponding segmentation masks. To ensure the fairness and statistical significance of model performance evaluation, the dataset was divided into training, testing, and validation sets at a ratio of 7:2:1.

Experimental setup

The training and evaluation of the MSDAF network were conducted on a Linux operating system, implemented using the PyTorch framework with a software environment consisting of CUDA 11.8, PyTorch 2.0.0, and Python 3.8. The hardware configuration for training included a 12 vCPU Intel(R) Xeon(R) Platinum 8352 V CPU @ 2.10 GHz and a 4090D graphics card. The Adam optimizer was uniformly adopted during the model training phase, with an initial learning rate set to 1e-3 and a weight decay coefficient configured as 1e-4. Weight decay was used to suppress overfitting of the model. Additionally, a gradient clipping mechanism was introduced to avoid gradient explosion during training, with the maximum gradient norm limited to 5.0.

Evaluation metrics

To comprehensively evaluate the performance of the model, multiple dimensions including pixel-level classification accuracy, regional overlap, and boundary fitting degree were considered. For this purpose, five widely recognized evaluation metrics in academia were selected, with their definitions and calculation formulas presented as follows:

Dice Similarity Coefficient [43]: The Dice coefficient is suitable for evaluating segmentation performance in class-imbalanced scenarios. It calculates the ratio of twice the intersection between the model-predicted segmentation region and the ground-truth annotated region to the sum of the pixel counts of the two regions.

16

where TP (True Positive) denotes the number of pixels correctly predicted as infected regions, FP (False Positive) denotes the number of pixels incorrectly predicted as infected regions from the background, FN (False Negative) denotes the number of pixels incorrectly predicted as background from infected regions, and TN (True Negative) denotes the number of pixels correctly predicted as background.

Intersection over Union (IoU) [44]: It calculates the ratio of the intersection to the union of the predicted region and the ground-truth region.

17

Precision: It measures the proportion of pixels truly positive among all pixels predicted as positive by the model.

18

Recall: It measures the proportion of truly positive pixels correctly predicted by the model among all truly positive pixels.

19

Accuracy: It calculates the proportion of correctly classified pixels among all pixels.

20

Baseline

To comprehensively validate the segmentation performance of MSDAFNet, a baseline model is constructed using MobileNetV3-large as the backbone within a basic U-Net architecture. This baseline model retains only the classic encoder-decoder structure and basic cross-level feature concatenation.

For comparison, fourteen representative segmentation networks in the medical image segmentation domain are selected as baseline models. These models cover different technical routes. Specifically, the classical CNN-based segmentation models include FCN [38], SegNet [45], UNet [15], UNet++ [14], HRNet [46], PAN [47], DeepLabV3+ [48], and PSPNet [49]. In addition, a Transformer-related lightweight hybrid segmentation network, MobileViT [50], is included to evaluate the performance of the proposed method against Transformer-based feature modeling. Furthermore, recent efficient segmentation networks involving lightweight convolutional design, attention-based decoding, and Mamba-style visual representation are also compared, including MKUNet [51], EMCADNet [52], MCADSDecoder [53], LVUNet [54], and VMUNet [55]. Among them, VMUNet is included as a representative Vision Mamba-based segmentation model, with Vision Mamba providing an efficient state-space visual representation paradigm [56].

To ensure a fair comparison, all models are trained and tested under identical experimental settings on four datasets: QaTa-COV19-v1, QaTa-COV19-v2, M-QaTa-COV19-v1, and M-QaTa-COV19-v2.

Results and analysis

Main results

Qdata-COVID19

The quantitative results on the QaTa-COV19-v1 and QaTa-COV19-v2 datasets are presented in Tables 1 and 2, respectively. MSDAFNet achieves competitive performance on the QaTa-COV19-v1 and QaTa-COV19-v2 datasets across six evaluation metrics, including Dice, IoU, Precision, Recall, F1, and Accuracy. Compared with representative segmentation networks such as FCN, HRNet, UNet++, DeepLabV3+, MobileViT, and VMUNet, the proposed method obtains consistently improved results in most metrics.

Table 1.

Results on Qdata-COVID19-v1

Model	Dice↑	IOU↑	Precision↑	Recall↑	F1↑	Accuracy↑
FCN	0.8091	0.6794	0.8025	0.8159	0.8091	0.9520
HRNet	0.8127	0.6845	0.8015	0.8243	0.8127	0.9452
PAN	0.8255	0.7029	0.8343	0.8169	0.8255	0.9502
DeepLabV3+	0.8265	0.7043	0.8411	0.8124	0.8265	0.9575
PSPNet	0.7861	0.6476	0.8089	0.7647	0.7861	0.9481
UNet	0.8234	0.6998	0.8228	0.8240	0.8234	0.9490
UNet++	0.8082	0.6782	0.8360	0.7822	0.8082	0.9537
SegNet	0.8162	0.6895	0.8124	0.8201	0.8162	0.9540
MKUNet	0.8118	0.6833	0.8145	0.8092	0.8118	0.94337
EMCADNet	0.8170	0.6906	0.8169	0.8171	0.8170	0.9450
MCADSDecoder	0.8066	0.6758	0.7809	0.8340	0.8066	0.9410
LVUNet	0.8517	0.7417	0.8177	0.8886	0.8517	0.9535
MobileViT	0.8341	0.7154	0.8072	0.8628	0.8341	0.9485
VMUNet	0.8116	0.6829	0.8099	0.8133	0.8116	0.9444
Ours	0.8544	0.7459	0.8443	0.8648	0.8569	0.9559

Table 2.

Results on Qdata-COVID19-v2

Model	Dice↑	IOU↑	Precision↑	Recall↑	F1↑	Accuracy↑
FCN	0.8609	0.7558	0.8608	0.8610	0.8584	0.9633
HRNet	0.8413	0.7261	0.8151	0.8692	0.8413	0.9558
PAN	0.8487	0.7371	0.8547	0.8427	0.8487	0.9595
DeepLabV3+	0.8682	0.7671	0.8486	0.8888	0.8682	0.9652
PSPNet	0.8384	0.7218	0.8315	0.8454	0.8384	0.9561
UNet	0.8486	0.7369	0.8359	0.8616	0.8486	0.9585
UNet++	0.8450	0.7316	0.8006	0.8946	0.8450	0.9558
SegNet	0.8584	0.77519	0.8534	0.8635	0.8584	0.9633
EMCADNet	0.8479	0.7360	0.8350	0.8612	0.8479	0.9602
MKUNet	0.8627	0.7586	0.8361	0.8515	0.8650	0.9635
MCADSDecoder	0.8202	0.6952	0.8366	0.8044	0.8202	0.9563
LVUNet	0.8695	0.7707	0.8452	0.8974	0.8705	0.9455
MobileViT	0.8600	0.7543	0.8299	0.8911	0.8600	0.9526
VMUNet	0.8253	0.7026	0.8119	0.8392	0.8253	0.9587
Ours	0.8755	0.7785	0.8756	0.9003	0.8755	0.9670

On the QaTa-COV19-v1 dataset, MSDAFNet achieves a Dice of 0.8544 and an IoU of 0.7459. Compared with the strong baseline models, the improvement is reflected not only in the Dice score but also in the IoU value, indicating more complete regional overlap between the predicted masks and the ground truth. On the larger QaTa-COV19-v2 dataset, MSDAFNet achieves a Dice of 0.8755, an IoU of 0.7785, and a Recall of 0.9003. These results suggest that the proposed framework maintains stable segmentation performance under different lesion distributions and dataset scales.

As shown in Table 3, MSDAFNet contains 6.96 M trainable parameters and requires 1.14 G FLOPs. The total memory usage is 89.43 MB, including 27.86 MB parameter memory. The average inference latency is 2.69 ms, corresponding to 371.43 FPS under the same evaluation protocol. Compared with commonly used segmentation models such as UNet, UNet++, DeepLabV3+, SegNet, PSPNet, and PAN, MSDAFNet has a smaller parameter size, lower FLOPs, and moderate memory usage, while maintaining competitive segmentation performance.The module-level analysis in Table 7 further shows that the encoder and decoder stages account for most of the computational cost. The bottleneck stage, where the global context-aware memory modeling is incorporated, contains 1.01 M parameters and accounts for 67.79 M FLOPs, corresponding to only 5.92% of the total FLOPs.

Table 3.

Computational complexity comparison of different segmentation models

Model	Params (M)	FLOPs (G)	Memory (MB)	Latency (ms)	FPS
UNet	29.61	48.23	406.92	3.15	317.01
UNet++	34.93	138.17	603.08	10.36	96.53
FCN	22.42	5.37	200.12	1.31	764.39
DeepLabV3+	43.04	19.59	382.54	2.40	416.23
SegNet	24.31	25.20	236.39	1.90	527.60
PSPNet	44.42	6.63	439.65	1.84	543.12
HRNet	3.31	2.56	48.84	3.11	321.77
PAN	19.94	46.97	342.60	3.40	293.86
EMCADNet	7.19	0.47	77.24	1.83	546.54
MK-UNet	0.31	0.31	29.45	2.82	354.85
LV-UNet	0.91	0.19	193.01	1.53	655.14
MSDAFNet (Ours)	6.96	1.14	89.43	2.69	371.43

Table 7.

Module-level parameter and FLOPs distribution of MSDAFNet

Module	Params	Params Ratio	FLOPs	FLOPs Ratio
Encoder	2.834 M	40.70%	291.43 M	25.46%
Bottleneck	1.010 M	14.50%	67.79 M	5.92%
Decoder Block 5	2.378 M	34.15%	159.70 M	13.95%
Decoder Block 4	581.38K	8.35%	152.93 M	13.36%
Decoder Block 3	132.31K	1.90%	139.79 M	12.21%
Decoder Block 2	35.16K	0.50%	149.55 M	13.07%
Decoder Block 1	9.83K	0.14%	169.08 M	14.77%
Total	6.964 M	100.00%	1.144 G	100.00%

M-Qdata-COVID19

To specifically evaluate the segmentation performance under small-lesion conditions, cases with lesion areas below 6% were selected from the original QaTa-COV19-v1 and QaTa-COV19-v2 datasets. Based on these selected cases, two derived small-lesion evaluation subsets, M-QaTa-COV19-v1 and M-QaTa-COV19-v2, were constructed with a split ratio of 7:2:1. The M-QaTa-COV19-v1 subset contains 1246 chest X-ray images with corresponding segmentation masks, while the M-QaTa-COV19-v2 subset contains 2571 chest X-ray images with corresponding segmentation masks. The heatmaps and lesion distribution statistics of the original datasets and the derived small-lesion subsets are shown in Figs. 6 and 7.

Fig. 6.

Heat maps of lesion distribution for the four datasets

Fig. 7.

Bar charts of lesion distribution for the four datasets

The QaTa-COV19-v1 and QaTa-COV19-v2 datasets show symmetric and high activation signals in the middle and lower parts of both lung fields. These signals reflect typical imaging characteristics of diffuse alveolar lesions with wide coverage. In contrast, activation signals in the derived small-lesion datasets M-QaTa-COV19-v1 and M-QaTa-COV19-v2 are locally distributed in the middle lung fields. Quantitative results are summarized in Tables 4 and 5.

Table 4.

Results on M-Qdata-COVID19-v1

Model	Dice↑	IOU↑	Precision↑	Recall↑	F1↑	Accuracy↑
FCN	0.6088	0.4376	0.5791	0.6417	0.6088	0.9593
HRNet	0.5850	0.4134	0.6106	0.5615	0.5850	0.9644
PAN	0.5999	0.4285	0.5320	0.6876	0.5999	0.9694
DeepLabV3+	0.5791	0.4076	0.5266	0.6432	0.5791	0.9688
PSPNet	0.5220	0.3532	0.5407	0.5047	0.5220	0.9692
UNet	0.5979	0.4265	0.5414	0.6677	0.5979	0.9557
UNet++	0.6188	0.4481	0.6148	0.6229	0.6088	0.9621
SegNet	0.5202	0.3515	0.4550	0.6072	0.5202	0.9626
EMCADNet	0.5582	0.3871	0.5031	0.6268	0.5582	0.9669
MKUNet	0.5546	0.3837	0.5262	0.5862	0.5546	0.9686
MCADSDecoder	0.6040	0.4330	0.5703	0.6418	0.6040	0.9599
LVUNet	0.6258	0.4445	0.5818	0.6319	0.6058	0.9715
MobileViT	0.5703	0.3989	0.5565	0.5848	0.5703	0.9706
VMUNet	0.6006	0.4299	0.5817	0.6407	0.6006	0.9624
Ours	0.6293	0.4591	0.5806	0.6482	0.6125	0.9726

Table 5.

Results on M-Qdata-COVID19-v2

Model	Dice↑	IOU↑	Precision↑	Recall↑	F1↑	Accuracy↑
FCN	0.6835	0.5192	0.6406	0.7327	0.6835	0.9682
HRNet	0.6907	0.5276	0.6821	0.6996	0.6907	0.9716
PAN	0.6934	0.5306	0.6681	0.7206	0.6934	0.9711
DeepLabV3+	0.7072	0.5470	0.6979	0.7167	0.7072	0.9731
PSPNet	0.6504	0.4819	0.6048	0.7034	0.6504	0.9657
UNet	0.6834	0.5191	0.6176	0.7648	0.6834	0.9679
UNet++	0.6995	0.5378	0.6794	0.7208	0.6995	0.9719
SegNet	0.6854	0.5214	0.6250	0.7587	0.6854	0.9684
EMCADNet	0.6765	0.5111	0.6834	0.6696	0.6765	0.9710
MKUNet	0.6882	0.5247	0.6770	0.6999	0.6882	0.9713
MCADSDecoder	0.6300	0.4599	0.6811	0.5861	0.6300	0.9678
LVUNet	0.7196	0.5620	0.7023	0.7377	0.7196	0.9739
MobileViT	0.6929	0.5302	0.6332	0.7652	0.6929	0.9639
VMUNet	0.6670	0.5004	0.6114	0.7337	0.6670	0.9660
Ours	0.7309	0.5759	0.7038	0.7601	0.7309	0.9746

The MSDAFNet maintains high robustness in small-lesion segmentation tasks. On the M-QaTa-COV19-v1 and M-QaTa-COV19-v2 datasets, Dice coefficients reach 0.6293 and 0.7309, respectively. Corresponding IoU values are 0.4591 and 0.5759, and accuracy values exceed 0.97. The proposed model is significantly superior to other comparative methods.These qualitative comparisons further illustrate the behavior of different models under small-lesion and blurry-boundary conditions. Compared with other methods, MSDAFNet shows more stable lesion localization and more complete mask prediction in regions with weak contrast, irregular contours, and anatomical background interference.

Ablation experiments

To further analyze the contribution of different module combinations in MSDAFNet, systematic ablation experiments were conducted on the QaTa-COV19-v1 dataset. A U-shaped encoder-decoder network built on MobileNetV3 was used as the baseline model. Based on this baseline, seven non-baseline variants were constructed by progressively combining the Global Context-aware Memory Module, the gated skip fusion strategy, and the Multi-Scale Deformable Convolutional Attention (MSDA) module. The segmentation performance of each variant was evaluated using six metrics, including Dice, IoU, Precision, Recall, F1-score, and Accuracy. The detailed results are presented in Table 6.

Table 6.

Ablation experiments with different module combinations on QaTa-COV19-v1

Model	Memory	Gated	MSDA	Dice↑	IoU↑	Precision↑	Recall↑	F1↑	Accuracy↑
Baseline	×	×	×	0.8272	0.7210	0.8054	0.8449	0.8272	0.9549
+ Memory	\checkmark	×	×	0.8345	0.7256	0.8168	0.8507	0.8345	0.9554
+ Gated	×	\checkmark	×	0.8329	0.7241	0.8136	0.8521	0.8329	0.9562
+ MSDA	×	×	\checkmark	0.8368	0.7287	0.8205	0.8542	0.8368	0.9551
+ Memory + Gated	\checkmark	\checkmark	×	0.8453	0.7321	0.8304	0.8609	0.8453	0.9527
+ Memory + MSDA	\checkmark	×	\checkmark	0.8386	0.7220	0.8237	0.8540	0.8386	0.9583
+ Gated + MSDA	×	\checkmark	\checkmark	0.8475	0.7354	0.8418	0.8533	0.8475	0.9539
MSDAFNet	\checkmark	\checkmark	\checkmark	0.8544	0.7459	0.8443	0.8648	0.8569	0.9559

As shown in Table 6, the baseline model achieves a Dice of 0.8272 and an IoU of 0.7210. After introducing a single component, the segmentation performance improves to different degrees. The model with only the Memory Module obtains a Dice of 0.8345, indicating that global contextual guidance is beneficial for improving lesion-region discrimination. The model with only the gated skip fusion strategy achieves a Dice of 0.8329, suggesting that cross-level feature calibration contributes to the preservation of lesion-related spatial details. The model with only the MSDA module reaches a Dice of 0.8368 and an IoU of 0.7287, showing that multi-scale deformable feature aggregation is helpful for representing lesions with scale variation and irregular contours.

When two modules are combined, the performance is further improved in most cases. The combination of the Memory Module and gated skip fusion achieves a Dice of 0.8453 and a Recall of 0.8609, indicating that global semantic guidance and boundary-aware feature calibration can jointly enhance lesion recall. The combination of the Memory Module and MSDA obtains a Dice of 0.8386, while the combination of gated skip fusion and MSDA reaches a Dice of 0.8475 and an IoU of 0.7354. These results suggest that MSDA and gated skip fusion are complementary in improving boundary-sensitive feature representation and cross-level feature selection.

The complete MSDAFNet, which integrates all three components, achieves the highest Dice, IoU, Precision, Recall, and F1-score among the evaluated variants, with values of 0.8544, 0.7459, 0.8443, 0.8648, and 0.8569, respectively. Although the Accuracy values of different variants remain close due to the large proportion of background pixels in CXR segmentation, the complete model shows the most balanced performance in terms of region overlap and lesion recall.

Computational complexity analysis

To further evaluate the computational characteristics of the proposed MSDAFNet, we report the number of parameters, FLOPs, memory usage, inference latency, and FPS under the same evaluation setting. The results are summarized in Table 3. In addition, to better understand the computational distribution within MSDAFNet, a module-level complexity analysis is provided in Table 7.

To provide an intuitive interpretation of the ablation results, Grad-CAM heatmaps of three representative cases are presented in Fig. 8. In this figure, Base denotes the baseline encoder-decoder model, A denotes the variant equipped with the Global Context-aware Memory Module, B denotes the variant incorporating both the Memory Module and the gated skip fusion strategy, and C denotes the final model with the additional MSDA module.

Fig. 8.

Grad-CAM-based visualization of progressive module effects. Base denotes the baseline encoder-decoder model. A denotes the model after adding the global context-aware memory module, B denotes the model after further adding the gated skip fusion strategy, and C denotes the final model after introducing the MSDA module

In Case 1, the baseline model can roughly activate the main lesion region, but it also produces a relatively strong response outside the annotated mask. After introducing the Memory Module, the activation becomes more concentrated around lesion-related regions, indicating improved global semantic discrimination. With the further incorporation of gated skip fusion, the response along the lesion boundary becomes more continuous, suggesting that lesion-related spatial details are better preserved during cross-level feature fusion. The final model further suppresses irrelevant activation and produces a more focused response over the lesion area. In Case 2, where bilateral diffuse lesions are present, the baseline model only highlights partial lesion regions. The activation gradually expands after adding the Memory Module and gated skip fusion, while the final model shows better spatial consistency with the large lesion mask. In Case 3, anatomical interference leads to scattered activation in non-lesion areas. The final model effectively reduces these irrelevant responses and retains activation mainly around the annotated bilateral lesions.

These qualitative observations are in accordance with the quantitative trends reported in Table 6. With the progressive introduction of the proposed components, the activation maps show a clearer concentration on lesion-relevant regions and reduced responses to surrounding anatomical structures. This evolution of activation patterns is consistent with the improvements in Dice, IoU, Recall, and F1-score, indicating enhanced region overlap and lesion recall. Overall, the Grad-CAM results provide complementary visual evidence that the proposed components contribute synergistically to lesion localization, boundary continuity, and mask completeness.

Case study

Analysis on the Qdata-COVID19 dataset

To intuitively present the performance differences of various segmentation methods, eight representative images are selected from the dataset. These images cover typical clinical scenarios with multiple common lesion types, as shown in Fig. 9. Consistent with trends shown by quantitative experimental results, other networks generally suffer from missed lesion detection and mis-segmentation of edge regions in complex scenes. Effective segmentation is usually achieved only when the contrast between lesions and the background is high and structural interference is low. As shown in Case 1 of Fig. 10, most models can identify the general lesion region when the lesion area is large and occlusion interference is limited. Few false detections are observed in other background regions. However, in Case 2 of Fig. 10, even with a large lesion area, models that performed well in Case 1, such as MOENet and MKUNet, show large-area missed detection under indistinct edge transitions and severe internal noise. In contrast, the MSDAFNet establishes better global context relations through the memory module. The adaptability of the model to individual differences in anatomical structures is enhanced. Better robustness is achieved in distinguishing targets from background noise. Similarly, in Cases 3 and 4 of Fig. 10, traditional methods often misclassify mediastinal tissues and vascular shadows as lesions, or produce rough segmentation contours due to insufficient edge refinement. With its discriminative feature extraction scheme and more flexible convolution structure, MSDAFNet effectively covers the scale range from tiny nodules to diffuse lesions. Accurate fitting of blurred boundaries is realized when facing irregular lesions and indistinct borders between lesions and normal lung tissues. In addition, traditional methods are prone to large-area false detection in fine small-lesion segmentation tasks. This issue is caused by minimal gray differences in surrounding lung parenchyma and significant background interference from fine blood vessels. In Case 5 of Fig. 10, most models except DeeplabV3+ and MOENet show false detections in the left noise region for small-lesion segmentation in low-quality chest X-ray images. This situation becomes more obvious in Case 6 of Fig. 10, where the gray difference between lesions and lung parenchyma is insignificant. In comparison, MSDAFNet adjusts sampling positions through deformable convolution to fit occluded lesion edges. The segmented small white lesions are complete with clear boundaries, free of large-area false-positive segmentation caused by overlapping structures. In Case 7 of Fig. 10, feature signals of small lesions are further weakened by noise, artifacts, and uneven image contrast. Most models misclassify tissue interference in the left lower region as lesions due to the inability to distinguish noise from weak lesion features. MSDAFNet effectively filters invalid information corresponding to noise and tissue interference via gated attention, and accurately captures fine-grained features of small lesions.

Fig. 9.

Deformable convolution

Fig. 10.

Visual comparison results of different models

Analysis on M-Qdata-COVID19 dataset

To further evaluate the robustness of the proposed network under small-lesion conditions, segmentation performance of all networks is tested on the derived small-lesion dataset, as illustrated in Fig. 11.

Fig. 11.

Visual comparison of different models on the small-lesion dataset

It can be observed that MSDAFNet maintains better robustness even with a smaller dataset. This improvement is attributed to the cooperative framework of the memory module and gated attention.

In Case A of Fig. 11, most models show large-area missed detection for the lesion region in the upper-left lung. In contrast, MSDAFNet accurately segments the upper-left lesion and completely covers all lesions in the lower-right region.

In Cases B and C of Fig. 11, EMCADNet and MobileViT, which performed well previously, lose robustness. Various degrees of false detection are observed under the interference of ribs and lung parenchyma.

These visualization results are highly consistent with quantitative experimental performance. The effectiveness and stability of MSDAFNet for accurate chest X-ray lesion segmentation in complex clinical scenarios are fully verified.

Discussion

The experiments are mainly conducted on QaTa-COV19-v1, QaTa-COV19-v2, and their derived small-lesion subsets. These datasets provide useful pixel-level annotations for evaluating chest X-ray lesion segmentation, but they are still derived from the same dataset series. In practical clinical scenarios, chest X-ray images may vary across institutions due to differences in imaging protocols, equipment settings, patient populations, disease categories, image quality, and annotation criteria. Such variations may influence the transferability of segmentation models across different data sources.

Therefore, external validation on additional chest X-ray segmentation datasets with dense pixel-level annotations and multi-center clinical data remains an important direction for future work. In addition, future studies may incorporate boundary-sensitive metrics, such as HD95, ASSD, or Boundary F1, to more directly evaluate boundary delineation quality. More fine-grained analyses of small lesions, low-contrast lesions, and cases with severe anatomical interference would also be valuable for further assessing the robustness and clinical applicability of the proposed method.

Conclusion

To realize the accurate segmentation of pathological structures in chest X-ray images, this study proposes a multi-scale deformable convolutional attention fusion segmentation network, MSDAFNet. A complete segmentation processing pipeline is built on the network architecture, which synergistically integrates three core modules: the multi-scale deformable convolutional attention module, the gated attention mechanism and the global context-aware memory module. This integrated architecture innovatively tackles the key challenges in chest X-ray segmentation, including the structural complexity and blurry boundaries of lesion regions. Extensive experiments are conducted on two public datasets, QaTa-COV19-v1 and QaTa-COV19-v2, to verify the effectiveness and robustness of the proposed network. The network achieves Dice coefficients of 0.8544 and 0.8755 on the two datasets respectively, with Intersection over Union values of 0.7459 and 0.7785. These results demonstrate the competitive performance of MSDAFNet on the corresponding segmentation tasks, providing a promising solution for accurate chest X-ray pathological structure segmentation.

Author contributions

X.L. wrote the main manuscript ; J.Z. processed research data; Z.L. developed research methods; Y.Y. analyzed research data; Y.L. assisted with manuscript drafting; Y.X. prepared figures and tables; X.H. supervised the research and revised the manuscript; all authors reviewed the final version of the manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (No. 62262029), the Natural Science Foundation of Jiangxi Province (No. 20212BAB202016).

Data availability

All data used and/or analysed during this study are included in the published articles entitled COVID-19 infection map generation and detection from chest X-ray images and Reliable covid-19 detection using chest x-ray images [and their supplementary information files].

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.