基于DCNN的椎体分割与骨密度计算方法研究

Research on Vertebral Segmentation and Bone Mineral Density Calculation Method Based on DCNN

摘要: 本研究旨在探讨深度学习在原发性骨质疏松症患者中的应用。通过采用基于深度卷积神经网络(DCNN)的全自动方法，实现椎体分割和CT图像中的骨密度(BMD)计算。采用588例患者作为训练数据，以及863例患者作为测试数据。首先使用U-Net全卷积神经网络对椎体进行自动分割，并以手动勾画的椎体区域作为对照。其次采用DenseNet-121卷积神经网络对BMD进行计算，并与定量CT (QCT)的标准值进行比较。测试集分为三组：测试集1 (463例)、测试集2 (200例)和测试集3 (200例)。实验结果显示自动分割结果与手动分割结果高度相关，三个测试集的Dice系数分别可达0.823、0.786和0.782。不同供应商的测试集显示，自动计算的平均BMD值与QCT结果高度相关(相关系数r > 0.98)。最终研究结论表明，基于深度学习的方法能够实现CT图像中骨质疏松症、骨量减少和正常骨密度的全自动识别，具有良好的精确度和一致性。

Abstract: The study aims to explore the application of deep learning in patients with primary osteoporosis. By adopting a fully automated method based on Deep Convolutional Neural Networks (DCNN), vertebral segmentation and Bone Mineral Density (BMD) calculation from CT images are achieved. A total of 588 patients were used for training data, and 863 patients were used for testing data. First, a U-Net fully convolutional neural network was employed to automatically segment the vertebrae, with manually drawn vertebral regions as a reference. Next, the DenseNet-121 convolutional neural network was used to calculate BMD and compare it with the standard values of quantitative CT (QCT). The test set was divided into three groups: Test Set 1 (463 cases), Test Set 2 (200 cases), and Test Set 3 (200 cases). Experimental results showed that the automatic segmentation results were highly correlated with manual segmentation, with Dice coefficients of 0.823, 0.786, and 0.782 for the three test sets, respectively. The test sets from different vendors showed that the automatically calculated average BMD values were highly correlated with the QCT results (correlation coefficient r > 0.98). The final conclusion of the study indicates that the deep learning-based method can achieve fully automated identification of osteoporosis, osteopenia, and normal bone density in CT images, with good accuracy and consistency.