In the field of natural images, Li et al. [22] used a dual attention fusion module to effectively combine features from images from different sources, thereby enhancing the model’s ability to focus on important regions. The attention mechanism can enhance the extraction of key features in infrared and visible images, making the fused images clearer and retaining more meaningful details [23]. In medical image segmentation, the attention mechanism is primarily used to guide the model’s focus on the lesion areas in the image, assigning different weights to each pixel or feature, enhancing task-relevant features, and suppressing irrelevant background information. Huang et al. [24] prior convolutional attention mechanism that dynamically allocates attention weights across both channel and spatial dimensions. Shaker et al. [25] used a pair of mutually dependent branches based on spatial and channel attention to effectively learn discriminative features, improving the quality of segmentation masks. Fu et al. [26] used a Transformer-based spatial and channel attention module to extract global complementary information across different layers of the U-Net, which helps in learning detailed features at different scales. To address hair interference in dermoscopic images, Xiong et al. [27] proposed a multi-scale channel attention mechanism that enhances feature information and boundary awareness. Song et al. [28] argued that current popular attention mechanisms focus too much on external image features and lack research on latent features. They introduced an external-latent attention mechanism, using an entropy quantization method to summarize the distribution of latent contextual information. Similarly, Huang et al. [29] used Bi-Level Routing Attention in deep networks to discard irrelevant key-value pairs, achieving content-aware sparse attention for dispersed semantic information.

While pursuing high performance, researchers have also begun to focus on the lightweight and efficiency of medical image segmentation networks. Network structure design is one of the most popular approaches for lightweight optimization. Ma et al. [17] simplified the structure, reduced the number of parameters, and optimized the convolution operations, achieving a significant reduction in computational complexity and model size while maintaining segmentation accuracy. This enables the model to perform excellently even in resource-constrained environments, making it suitable for applications such as mobile healthcare and telemedicine. The UcUNet [30] network achieves lightweight and precise medical image segmentation by designing an efficient large-kernel U-shaped convolution module. This network leverages large-kernel convolutions to expand the receptive field while integrating depthwise separable convolutions to reduce the computational cost, thereby maintaining high segmentation accuracy with efficient computation. Liu et al. [16] combines the lightweight characteristics of HarDNet with multi-attention mechanisms, enhancing the network’s ability to capture key features and achieving more precise medical image segmentation. Sun et al. [31] introduces a contextual residual network, effectively integrating contextual information into the U-shaped network, enhancing the global understanding and stability of the segmentation. Nisa and Ismail [32] employs a dual-path structure with a ResNet encoder, combining ResNet’s feature extraction capabilities with U-Net’s segmentation advantages, offering an alternative effective solution for medical image segmentation. Zhao et al. [33] proposed a four-layer feature calibration branch based on an attention mechanism. The downsampling layer reduces the resolution of rectal cancer CT image feature maps to half of the original size, followed by pointwise convolution to enable interactions between channels. This method effectively expands the receptive field of subsequent convolutional layers and optimizes computational efficiency by reducing the cost of calculating spatial attention. Model compression, as another approach to simplifying network structures, removes structural redundancy while maintaining performance, making it more suitable for various applications in medical image analysis. Wang et al. [34] designed a sophisticated teacher network to learn multi-scale features, guiding a more lightweight student network to improve segmentation accuracy. Experiments showed that this method effectively acquires detailed morphological features of the brain from the teacher network. Hajabdollahi et al. [35] proposed a channel pruning algorithm for medical image segmentation tasks, which selects color channels during image processing and allows training of the target structure directly on the pre-selected key channels. However, these studies did not address how to utilize the redundancy effectively.

Based on the above research findings, this paper proposes a lightweight segmentation model that emphasizes spatial and channel features. This model improves segmentation accuracy and efficiency without increasing additional computational costs, providing a new and efficient solution for the medical imaging field.

As illustrated in Figure 1, the proposed model framework consists primarily of the ISDConv module, the AEAM module, and dynamic convolution. The ISDConv module is composed of three parts: ScConv, Inception convolution, and standard convolution. By incorporating depthwise separable convolutions and group convolutions, ISDConv facilitates the model’s understanding of multi-scale information within images, thereby enhancing its ability to detect and classify objects of varying sizes.

The AEAM module operates in two stages: SEattention and SCSA. SEattention enhances the network’s representational capacity by explicitly modeling the interdependencies between convolutional feature channels. SCSA, in turn, is divided into two components: SMSA and PCSA. SMSA integrates multi-semantic information and employs a progressive compression strategy to inject discriminative spatial priors into the channel self-attention mechanism of PCSA, effectively guiding channel recalibration. Within PCSA, robust feature interaction based on a self-attention mechanism further mitigates the multi-semantic information discrepancy among sub-features in SMSA.

As shown in Figure 1, ISDConv consists of a ScConv, an Inception Convolution, and a standard Conv2d layer. The Inception Convolution achieves lightweight performance by efficiently decomposing a large kernel depthwise convolution into four parallel branches along the channel dimension. These branches consist of a small square kernel, two orthogonal large kernels, and an identity mapping. The use of a small square kernel reduces computational complexity, while the orthogonal large kernels capture different spatial information at varying scales. The identity mapping helps preserve the original input features, further enhancing the efficiency of the network. Additionally, this architecture incorporates 1 × 1 convolutions for dimensionality reduction before applying computationally expensive operations, minimizing the computational burden while preserving the model’s ability to learn rich, multi-scale features. These four branches not only achieve higher computational efficiency than the large kernel depthwise convolution but also maintain a large receptive field, enabling the model to capture spatial context effectively for improved performance.

One of the branches employs a 3 × 3 kernel, which avoids the inefficiency of large square kernels. Instead, large square kernels k h × k w are decomposed into 1 × k w and k h × 1 , significantly reducing computational complexity. Specifically, for a given input x , it is divided into four groups along the channel dimension, with the operation defined as Equation 1: X h w , X w , X h , X i d = S p l i t X = X : , : , g , X : g : 2 g , X : 2 g : 3 g , X : 3 g : (1) where, g represents the number of channels in each convolution branch, which is determined by the formula g = r g C , where r g is the ratio for splitting and C is the total number of input channels. The input is divided into four groups along the channel dimension based on this ratio, and the resulting split inputs are then fed into the respective parallel branches. Therefore, the following Equation 2 can be established: X h w ′ = D W C o n v k s × k s g → g g X h w X w ′ = D W C o n v 1 × k b g → g g X w X h ′ = D W C o n v k b × 1 g → g g X h X i d ′ = X i d (2) where k s represents the 3 × 3 kernel size, k b denotes the kernel sizes of 11 × 1 and 1 × 11 , X h w represents the feature map, X w refers to the features in the width direction, and X h refers to the features in the height dimension of the image. After processing each input x i through its respective branch, the outputs X ′ are concatenated along the channel dimension. The operation can be expressed as Equation 3. X ′ = C o n c a t X h w ′ , X w ′ , X h ′ , X i d ′ (3)

This paper introduces the AEAM attention module, designed to achieve synergy between channel attention and spatial attention through dimensional decoupling, lightweight multi-semantic guidance, and semantic discrepancy mitigation. As shown in Figure 2, the AEAM module consists of two main components: SEattention and SCSA.

The SCSA module is composed of two sequentially linked components: Shared Multi-Semantic Spatial Attention (SMSA) and Progressive Channel Self-Attention (PCSA). SMSA employs multi-scale, depth-sharing one-dimensional convolutions to extract spatial information at different semantic levels from four independent sub-features. This approach enables the efficient integration of diverse spatial semantics across sub-features. After SMSA modulates the feature maps, the resulting features are passed to PCSA. This component combines a progressive compression strategy with a channel-specific self-attention mechanism (CSA) to refine the feature representation further.

In this paper, a given input X ∈ R B × C × H × W is applied global average pooling along the height and width dimensions to create two unidirectional 1D sequence structures: X H ∈ R B × C × W , X W ∈ R B × C × H . To learn diverse spatial distributions and contextual relationships, the feature set is divided into K equally sized and independent sub-features, such that X H i and X W i , each sub-feature has a channel count of C K , where C is the total number of channels in the original feature set. In this study, we set the default value K = 4 , decomposing the features into H -dimensional and W -dimensional sub-features. During the decomposition process, 1D convolution is applied to each sub-feature. We employ lightweight shared convolutions for alignment, which implicitly model feature consistency across both dimensions by learning correlations.

The ablation formula is shown in Equation 4: X ̃ H i = D W C o n v 1 d k i C K → C K X H i X ̃ W i = D W C o n v 1 d k i C K → C K X W i (4) Where X H and X W represent feature maps in height and width dimensions respectively. SEattention introduces the “Squeeze-and-Excitation” (SE) block, which enhances the network’s representational capacity by explicitly modeling the interdependencies between convolutional feature channels. The SE block employs a special mechanism that enables the network to perform feature recalibration. Through this mechanism, the block learns to selectively emphasize informative features while suppressing less useful ones by leveraging global information.

The structure of the SE block is illustrated in the lower part of Figure 2. For any given transformation F t r , which maps the input X to a feature map U , which U ∈ R H × W × C , a corresponding SE block can be constructed to perform feature recalibration. The feature map U first undergoes a squeeze operation, which aggregates the feature map across the spatial dimensions to generate a channel descriptor. The function of this descriptor is to embed the global distribution of channel feature responses, thereby enabling all layers of the network to utilize information from the global receptive field. After the aggregation, an excitation operation follows. This operation, in the form of a simple self-gating mechanism, takes the embedding as input and generates a set of modulation weights for each channel. These weights are applied to the feature map U to produce the output of the SE block, which can then be directly fed into subsequent layers of the network.

In this study, each image in the dataset is associated with a corresponding binary mask. Skin lesion segmentation is treated as a pixel-level binary classification task, distinguishing the skin lesions from the background. The combination of Binary Cross-Entropy (BCE) loss and the Dice Similarity Coefficient (DSC) loss is used as the loss function to optimize the network parameters. This approach effectively addresses the challenge of skin lesion segmentation by balancing pixel accuracy and overlap between the predicted and ground truth masks.

The loss function, referred to as the BceDice loss, can be expressed as Equation 5: L BCE = − 1 N ∑ i = 1 N y i l o g p i + 1 − y i l o g 1 − p i L Dice = 1 − 2 | X ∩ Y | | X | + | Y | L BCEDice = α 1 L BCE + α 2 L Dice (5) where N is the total number of samples, Y represents the ground truth label, p i represents the predicted values, y i denotes the true label of sample i . | X | and | Y | denote the ground truth and the intersection of the predicted region, respectively. α 1 and α 2 represent the weights of the two loss functions. In this study, both weights are set to 1 by default.

The ISIC (International Skin Imaging Collaboration) datasets are benchmark datasets widely used in medical image analysis, particularly for dermoscopic image segmentation, classification, and automated skin cancer detection. These datasets feature high-resolution dermoscopic images with comprehensive annotations, including lesion boundaries, diagnostic labels, and metadata. Covering a diverse range of skin conditions, they are designed to support tasks such as lesion segmentation, feature extraction, and disease classification. Notably, the ISIC2017 and ISIC2018 datasets have been instrumental in advancing research on melanoma detection and other skin diseases through the annual ISIC Challenges. Our research is specifically conducted on the ISIC2017 and ISIC2018 datasets. Figure 3 are some sample images from the ISIC2017 and ISIC2018 datasets.

All experiments were implemented using the PyTorch framework and performed on a laptop equipped with an NVIDIA GeForce RTX 3080 Ti GPU with 8 GB of memory. Based on established practices, all images were normalized and resized to 256 × 256 pixels. Data augmentation techniques, including vertical flipping, horizontal flipping, and random rotations, were applied. The loss function used was the BCE-Dice loss, as defined in Equation 6. L BCE−Dice = α ⋅ − 1 N ∑ i = 1 N y i ⋅ l o g y i ̂ + 1 − y i ⋅ l o g 1 − y i ̂ + β 1 − 2 ⋅ ∑ i = 1 N y i ⋅ y i ̂ + ϵ ∑ i = 1 N y i + ∑ i = 1 N y i ̂ + ϵ (6) where y i represents the ground truth label, y i ̂ denotes the predicted value, N is the total number of pixels, ϵ is a small constant which is set to 10 in this work, α and β are the weights for the BCE and Dice components. AdamW was utilized as the optimizer with an initial learning rate of 0.001, dynamically adjusted using a cosine annealing scheduler. The maximum number of iterations was set to 50, with a minimum learning rate of 0.0001. The training process was conducted over 300 epochs with a batch size of 8.

In this study, segmentation performance is assessed using the mean Intersection over Union (mIoU), Dice Similarity Coefficient (DSC), and Accuracy (Acc), as defined in Equation 7. Additionally, the number of parameters is represented by Params, measured in millions (M), and computational complexity is quantified in GFLOPs. It is important to note that both Params and GFLOPs are calculated based on an input size of 256 × 256 . m I o U = T P T P + F P + F N D S C = 2 T P 2 T P + F P + F N (7) Where, TP, FP, FN, and TN represent True Positives, False Positives, False Negatives, and True Negatives, respectively.

In this section, we conducted comparative experiments on melanoma segmentation using the ISIC2017 and ISIC2018 skin lesion segmentation datasets and evaluated the test results. The evaluation metrics include DSC, mIoU, params, and GFLOPs. The results are presented in Tables 1, 2, where we perform a comprehensive comparison of the proposed model with the following methods: UNet [36], Transfuse [37], FATNet [38], MALUNet [39], QGD-Net [40], LCA-UNet [41], SCSONet [42], PL-Net [43], UCM-Net [44], CSAP-UNet-S [45], and ELA-Net [46].

In addition, bar charts are utilized in this study to visually illustrate the performance of different models on various metrics, providing a clearer comparison between our method and others. Specifically, for the comparison of lightweight metrics, only models designed with lightweight objectives were selected, with the results presented in Figures 4, 5. The experimental results indicate that MCANet outperforms all other methods in both data sets in terms of DSC and mIoU metrics. Notably, MCANet achieves Dice scores exceeding 0.9 on the ISIC datasets, significantly outperforming all comparison models and demonstrating its superior segmentation performance.

Furthermore, to further validate the segmentation performance of the model, we present the visual segmentation results on the ISIC dataset, as shown in Figures 6, 7. Although there are some differences between the MCANet segmentation results and mask images, MCANet outperforms other models in capturing detailed information from medical images, giving it a significant advantage in accurately segmenting the areas of the injury. Specifically, Figure 8 shows that MCANet can more accurately capture the target location in segmentation tasks involving smaller lesions, with finer and more precise segmentation of the lesion boundaries.However, our study also has some limitations. First, although MCANet demonstrates impressive performance on the ISIC datasets, its generalizability to other medical imaging datasets remains to be fully explored. In addition, while the model is lightweight in design, further optimization is required to meet the strict deployment constraints of resource-constrained devices, such as smartphones or embedded systems. Another limitation lies in the annotation quality of the datasets used, as potential noise in the segmentation masks may influence the model’s learning process. Finally, despite MCANet’s ability to capture detailed features, there are still some challenges in handling highly irregular or extremely small lesions, which may require more advanced attention mechanisms.To address these issues, future research will focus on several directions. First, extending the evaluation to additional datasets with diverse imaging modalities can help assess the robustness and versatility of MCANet. Second, incorporating techniques such as knowledge distillation or pruning could further improve the model’s efficiency for deployment in real-time scenarios. Third, exploring semi-supervised or unsupervised learning methods may reduce dependency on high-quality annotations, enabling better performance even with noisy labels. Finally, integrating advanced multi-scale feature extraction modules could enhance the model’s ability to handle challenging segmentation tasks involving complex lesion patterns.

To evaluate the contribution of each module in MCANet, we designed and conducted a series of ablation studies, with the results summarized in Table 3. Using the SCSONet baseline model as a reference, we performed comparative experiments with different combinations of the proposed modules on the ISIC dataset. Furthermore, to provide a clearer visualization of the impact of each module on segmentation performance, we used bar charts to illustrate variations in key metrics, such as DSC and mIoU, as shown in Figure 9.In the ablation study, “Base + AEAM” represents the integration of the proposed AEAM module into the baseline model, “Base + ISDConv” denotes the addition of the ISDConv module to the baseline, and “MCANet” refers to the complete network architecture proposed in this study. From Table 3 and the bar chart, it can be observed that integrating the proposed modules into the baseline model not only results in negligible increases in parameter count and computational complexity but also leads to significant improvements in segmentation performance. Specifically, as the modules are progressively added, the segmentation performance steadily improves, with the key metrics DSC and mIoU ultimately reaching 0.9086 and 0.8325, representing increases of 2.93 % and 5.37 % , respectively, compared to the baseline. The bar chart further illustrates this performance improvement trend, visually highlighting the contribution of each module.

Moreover, the experimental results demonstrate that the proposed modules collaborate effectively, with the addition of individual modules not causing any degradation in overall performance but instead continuously improving segmentation accuracy. Additionally, our module design is highly adaptable, allowing for seamless integration into other network architectures without requiring significant modifications to the original structure. For instance, incorporating the AEAM or ISDConv modules into other networks results in varying degrees of performance improvement, validating the generalizability and practicality of the proposed modules.

In summary, the results of the ablation studies and their visual analysis demonstrate the significant contributions of the proposed modules to the model’s performance. These improvements not only enhance the segmentation capability of MCANet but also highlight the academic significance and practical applicability of our work in the field of medical image segmentation.