摘要： 头影标志点检测对于临床诊断、治疗计划和研究至关重要。为提高自动检测的准确性，提出一种改进的多尺度特征融合检测模型AIW-Net。采用经过预训练的轻量型网络MobileNetV2作为主干网络进行特征提取，使用上、下采样路径构成中间模块，利用改进的倒残差模块减少下采样过程中的特征损失。在解码器模块中引入从粗到细的中间监督，将得到的多个尺度热图与特征图进行融合，并在跳跃连接中使用注意力门，有效抑制特征图中的背景区域响应。在ISBI 2015 Grand Challenge提供的基准数据集Test 1上进行实验，结果表明，该模型的平均径向误差为1.14 mm，在临床可接受的误差范围2 mm与2.5 mm内的成功检测率分别为86.38%与92.10%，性能优于W-Net、IW-Net等模型。

关键词: 标志点检测, 多尺度特征融合, 倒残差结构, 注意力门, 热图回归

Abstract: The detection of cephalometric mark point is extremely crucial for clinical diagnosis, treatment planning, and research.To improve the accuracy of automatic detection, an improved multi-scale feature fusion detection model, AIW-Net, is proposed.AIW-Net uses the pre-trained lightweight network MobileNetV2 as the backbone network for feature extraction. The intermediate module comprises up and down sampling paths, and the improved inverse residual module is used to reduce the feature loss in the down-sampling process.The intermediate supervision from coarse to fine is introduced into the decoder module, the obtained multi-scale heat map is fused with the feature map, and the attention gate is used in the jump connection to effectively suppress the response of the background region in the feature map.The experimental results of the benchmark dataset Test 1 provided by ISBI 2015 Grand Challenge show that the proposed model achieves a Mean Radial Error(MRE) of 1.14 mm, and a Successful Detection Rate(SDR) of 86.38% and 92.10% within the clinically acceptable error range of 2 mm and 2.5 mm, respectively, which are better than that obtained by W-Net, IW-Net and other models.

Key words: mark point detection, multi-scale feature fusion, inverse residual structure, attention gate, heatmap regression

中图分类号: 

• TP391.41