MRFA-Net：一种基于多尺度残差特征挖掘的肺癌CT图像分割网络-《中国医学物理学杂志》

文章信息/Info

Title:: MRFA-Net: a lung cancer CT image segmentation network based on multi-scale residual feature mining

作者:: 张奔腾1; 王娆芬1; 胡凌燕1; 王海玲1; 宫晓梅2; 1.上海工程技术大学电子电气工程学院，上海 201620；2.同济大学附属上海市肺科医院放疗科，上海 200433

Author(s):: ZHANG Benteng1; WANG Raofen1; HU Lingyan1; WANG Hailing1; GONG Xiaomei2; 1. School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China 2. Department of Radiation Therapy, Shanghai Pulmonary Hospital Affiliated to Tongji University, Shanghai 200433, China

Keywords:: Keywords: lung tumor segmentation residual feature extraction deep pyramid aggregation reverse attention mechanism

摘要:: 医学影像中的肺部肿瘤分割对于辅助诊断和治疗规划具有重要意义。然而，由于肿瘤在形态、边界和大小上的高度异质性，精准分割仍然具有较大挑战。提出一种基于残差特征挖掘与深层特征融合的深度分割网络MRFA-Net，旨在提升肺肿瘤在CT图像中的自动分割性能。提出多尺度残差特征提取模块和深层特征聚合模块，通过在多尺度上对残差特征进行提取以及深层特征融合，更有效地挖掘出具有代表性的肿瘤特征，并且在跳跃连接中引入反向注意力机制以提升模型对关键区域的感知能力。在MSD公开数据集和私有肺癌CT数据集上进行的实验结果表明MRFA-Net具备优异性能，Dice相似系数分别达到74.15%和74.77%。该方法为肺部肿瘤分割提供有效的解决方案，具备良好的临床应用潜力。

Abstract:: Abstract: Lung tumor segmentation from medical images is crucial for supporting diagnosis and treatment planning. However, tumors exhibit substantial heterogeneity in shape, contour and size, which makes precise segmentation still challenging. Therefore, this paper proposes a deep segmentation network named MRFA-Net (multi-scale residual feature aggregation network). Built upon residual feature mining and deep feature fusion, the network aims to improve the automatic segmentation performance of lung tumors on CT images. Specifically, two key modules are designed: the multi-scale residual feature extraction module and the deep pyramid aggregation module. By extracting residual features at multiple scales and fusing deep features, the network can capture representative tumor characteristics more effectively. Additionally, a reverse attention mechanism is integrated into skip connections to enhance the models perception of critical regions. Experimental results on the public MSD dataset and a private lung cancer CT dataset demonstrate that MRFA-Net achieves excellent performance, with a Dice similarity coefficient of 74.15% and 74.77%, respectively. The proposed method provides an effective solution for lung tumor segmentation and holds promising prospects for clinical practice.