Landmark localization with neural networks had gained popularity in recent years. However, due to the high dimensionality and large size of medical images, current neural network models still have problems such as information loss with deeper network, low accuracy and robustness. To address these issues, a 3D anatomical landmark localization method with a two-stage strategy was proposed in this study. The 3D spatial information between landmarks and the local information of each single feature point were extracted in these two stages. Additionally, new inception and attention modules were designed for the second stage to combine convolutional kernels of different sizes and weight labeling to strengthen the effective features extraction while weakening the invalid features. The proposed model was evaluated on a collected knee image dataset. The results outperformed state-of-the-art models with a mean error of 3.29 mm and a standard deviation of 2.17 mm. The outlier rates at error radius 3 mm, 5 mm and 7 mm were 53%, 22% and 5%, respectively, indicating good robustness of the model. The study provides a new neural network model with good accuracy for landmark localization tasks.

近年来，使用神经网络进行地标定位已经流行起来。然而，由于医学图像的高维和大尺寸，当前的神经网络模型仍然存在网络较深、精度低、鲁棒性低等问题。为了解决这些问题，本研究提出了一种具有两阶段策略的 3D 解剖标志定位方法。在这两个阶段中提取了地标之间的 3D 空间信息和每个单个特征点的局部信息。此外，为第二阶段设计了新的 inception 和 attention 模块，以结合不同大小的卷积核和权重标记，以加强有效特征提取，同时削弱无效特征。所提出的模型在收集的膝关节图像数据集上进行了评估。结果优于最先进的模型，平均误差为 3.29 毫米，标准偏差为 2.17 毫米。误差半径 3 mm、5 mm 和 7 mm 处的异常值率分别为 53%、22% 和 5%，表明模型具有良好的鲁棒性。该研究为地标定位任务提供了一种具有良好准确性的新神经网络模型。