利用深度学习构建自动分类模型以辅助髌骨轴位X线片的图像质控

Image quality control of axial patella X-ray images using automatic classification models based on deep learning

目的:利用深度学习方法训练髌骨轴位X线片图像质量控制的自动分类模型。方法:回顾性收集髌骨轴位X线片,由两位专家将髌骨轴位X线片分为不同数据组以训练模型,分别为:术后/非术后共175例(术后96例,非术后79例),侧别共735例(左侧419例,右侧316例),图像质量共453例(图像质量不合格246例,图像质量合格207例)。上述每组数据均按8:1:1的比例随机分为训练集、调优集和测试集,即:术后/非术后为136例、21例、18例,侧别586例、75例、74例,图像质量为362例、46例、45例。训练HRNet模型对上述三组图像进行自动分类,应用混淆矩阵评价模型分类预测效能,以符合率为评价指标。结果:测试集中,三组图像分类模型的预测符合率依次为:术后/非术后94.4%(17/18)、侧别98.6%(73/74)、图像质量91.1%(41/45)。结论:基于深度学习训练的分类模型对髌骨轴位X线片进行图像质量控制效能良好,有利于工作流程的优化及后续对接AI诊断模型。

Objective:To develop classification models based on deep learning to automatically classify the axial patella X-ray images in order to assist in image quality control.Methods:Axial patellar X-ray images were retrospectively collected.All images were reviewed by two radiologists and classified into three groups of binary classification,which included postoperative and nonoperative group(n=175,postoperative:n=96,nonoperative:n=79),laterality group(n=735,left:n=419,right:n=316),image quality group(n=453,poor image quality:n=246,good image quality:n=207).Each group data was randomly divided as train set,validation set and test set with the ration of 8:1:1.The classification models were trained using HRNet network,and the effectiveness of the models were evaluated using confusion matrix.Results:The accuracy of classification models of the axial patella X-ray images in the test set were as following:postoperative and nonoperative group 94.4%(17/18),laterality group 98.6%(73/74),and image quality group 91.1%(41/45).Conclusion:The models of automatic classification of axial patella X-ray images have a good performance,which was beneficial to the quality control and the subsequent implementation of AI diagnosis models.