<p>Medical image segmentation plays a critical role in diverse healthcare applications, providing essential support for early disease detection and comprehensive treatment planning. However, due to blurred boundaries, low contrast between lesions and surrounding tissues, and variations in lesion morphology and scale, existing segmentation methods still face significant challenges in enlarging the receptive field and extracting discriminative image features. To address these challenges, we propose MFA-Net, an innovative architecture specifically designed for accurate medical image segmentation. During the encoding stage, we employ convolutional kernel structures composed of 3<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>1 followed by 1<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>3 and 9<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>1 followed by 1<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>9 to construct the residual module, aiming to mitigate issues associated with the gradient vanishing and network performance degradation. In addition, we introduce a sophisticated channel-spatial attention mechanism in the last encoder to improve the capacity for feature extraction and representation. Furthermore, feature maps from adjacent encoder units are utilized to construct a multi-layer feature aggregation block, which allows features from different scales to synergistically contribute to the overall representation. Finally, to comprehensively evaluate the effectiveness and robustness of our approach, we conduct extensive experiments on two clinically significant jaw cyst datasets and the publicly available ISIC-2018 dataset. Comparative results demonstrate that MFA-Net outperforms current state-of-the-art methods across multiple evaluation metrics, achieving F1 of 0.9411, IoU of 0.9444, Mcc of 0.9406, and Jaccard of 0.8892 on the original jaw cyst dataset. Similarly, on the augmented jaw cyst dataset derived from the original jaw cyst images, MFA-Net maintains superior performance, achieving F1 of 0.9228, IoU of 0.9280, Mcc of 0.9221, and Jaccard of 0.8579. On the ISIC-2018 dataset, MFA-Net also achieves F1 of 0.8678, IoU of 0.8388, Mcc of 0.8238, and Jaccard of 0.7783. Moreover, the ablation study demonstrates the important contributions of the residual structure, channel-spatial attention mechanism, and multi-layer feature aggregation block to the overall performance of MFA-Net.</p>

Multi-layer feature aggregation network with residual module and attention mechanism for jaw cyst image segmentation

期刊论文
出版模式：

开放获取
发表日期：

2026年6月14日

查看全文

Huixia Zheng,
Xiaoliang Jiang,
Xu Xu

摘要

Medical image segmentation plays a critical role in diverse healthcare applications, providing essential support for early disease detection and comprehensive treatment planning. However, due to blurred boundaries, low contrast between lesions and surrounding tissues, and variations in lesion morphology and scale, existing segmentation methods still face significant challenges in enlarging the receptive field and extracting discriminative image features. To address these challenges, we propose MFA-Net, an innovative architecture specifically designed for accurate medical image segmentation. During the encoding stage, we employ convolutional kernel structures composed of 3 \(\times\) 1 followed by 1 \(\times\) 3 and 9 \(\times\) 1 followed by 1 \(\times\) 9 to construct the residual module, aiming to mitigate issues associated with the gradient vanishing and network performance degradation. In addition, we introduce a sophisticated channel-spatial attention mechanism in the last encoder to improve the capacity for feature extraction and representation. Furthermore, feature maps from adjacent encoder units are utilized to construct a multi-layer feature aggregation block, which allows features from different scales to synergistically contribute to the overall representation. Finally, to comprehensively evaluate the effectiveness and robustness of our approach, we conduct extensive experiments on two clinically significant jaw cyst datasets and the publicly available ISIC-2018 dataset. Comparative results demonstrate that MFA-Net outperforms current state-of-the-art methods across multiple evaluation metrics, achieving F1 of 0.9411, IoU of 0.9444, Mcc of 0.9406, and Jaccard of 0.8892 on the original jaw cyst dataset. Similarly, on the augmented jaw cyst dataset derived from the original jaw cyst images, MFA-Net maintains superior performance, achieving F1 of 0.9228, IoU of 0.9280, Mcc of 0.9221, and Jaccard of 0.8579. On the ISIC-2018 dataset, MFA-Net also achieves F1 of 0.8678, IoU of 0.8388, Mcc of 0.8238, and Jaccard of 0.7783. Moreover, the ablation study demonstrates the important contributions of the residual structure, channel-spatial attention mechanism, and multi-layer feature aggregation block to the overall performance of MFA-Net.