自动检测MR图像中的伪影：基于改进的AMAResNet网络

Automatic Detection of Artifacts in MR Images: An Enhanced AMAResNet-Based Network

摘要: 本文提出了一种创新网络结构，基于改进ResNet网络的算法，针对传统方法对局部特征和多尺度信息捕捉不足的问题，引入分块机制、Squeeze-and-Excitation (SE)模块和金字塔模块，分块机制将输入图像划分为局部子块以增强细节感知；SE模块通过通道权重自适应优化关键特征表达；金字塔结构融合跨尺度特征以提升复杂伪影识别能力。为解决类别不平衡问题，提出交叉熵与Dice损失的加权联合损失函数(最优权重比α = 0.7/β = 0.3)，构建出一个全新的名为AMAResNet的伪影检测模型，从而实现了MRI伪影的自动检测。该方法是数据驱动的，且不需要特殊的硬件要求，使用空间域中MR图像的大小作为输入。本文在真实数据上评估了算法性能，结果显示，改进后的网络在不同类型数据上的准确性和泛化能力方面均有显著提升。所提出的方法有望在扫描完成后立即将MR图像标记为无伪影或有伪影状态，从而确保获取高质量的MR图像，这对于准确的医学诊断至关重要。实验结果表明，改进后的ResNet网络在自动检测MRI伪影方面表现出色，具备在临床实践中广泛应用的潜力，有助于提高MRI图像质量和诊断准确性。

Abstract: This paper proposes an innovative network architecture based on an enhanced ResNet algorithm to address the limitations of traditional methods in capturing local features and multi-scale information. The framework introduces three key components: a partition mechanism that divides input images into local sub-regions to enhance detail perception, Squeeze-and-Excitation (SE) modules that adaptively optimize critical feature representations through channel-wise weighting, and a pyramid structure that integrates cross-scale features to improve complex artifact recognition. To resolve class imbalance issues, we design a weighted joint loss function combining cross-entropy and Dice Loss (optimal weight ratio α = 0.7/β = 0.3), ultimately constructing a novel artifact detection model named AMAResNet for automatic MRI artifact detection. This data-driven approach requires no specialized hardware and utilizes spatial domain MR image sizes as input. Evaluations of real-world datasets demonstrate that the enhanced network achieves significant improvements in both accuracy and generalization across different data types. The proposed method shows potential for immediate post-scan quality assessment by labeling MR images as artifact-free or artifact-affected, thereby ensuring high-quality medical imaging crucial for accurate diagnosis. Experimental results indicate that the improved ResNet architecture excels in automatic MRI artifact detection, exhibiting strong potential for clinical application to enhance both MRI image quality and diagnostic reliability.