基于卷积神经网络CT/CBCT影像质量自动分析

Automatic analysis of CT/CBCT image quality based on convolutional neural network

目的:通过独立的程序自动分析数据,可以在减轻影像的质量保证(QA)工作量的同时,尽可能避免操作者主观因素造成的偏差。方法:对Catphan500/503/504/600的CT/CBCT影像按照功能模块进行分类,并通过卷积神经网络(CNN)进行学习,学习后对新输入的CT/CBCT影像可以自动识别并根据功能模块进行分类,继而对相关指标包括影像CT值的线性、调制传递函数以及均匀性等进行自动分析,以便确保临床应用的影像质量达到要求。结果:对于Catphan500扫描的CT图像和Catphan503扫描的CBCT图像,经过CNN自动分类对于功能模块CTP401、CTP404、CTP528都可以正确标记出来,但是CTP486的精确度没有达到100%,即有部分不属于CTP486的模块被错误判断成CTP486。同时均可实现对CT的值线性、调制传递函数以及均匀性3个图像指标进行自动分析。结论:基于CNN能够准确地对CT/CBCT扫描的Catphan图像进行分类,下一步将拓展该方法到其他影像设备的QA体模,以便实现更广泛的自动影像质量保证。

Objective To propose an independent program for automatic data analysis which can avoid errors caused by subjective factors of the operator while reducing the quality assurance workload. Methods The computed tomography/cone beam computed tomography（CT/CBCT） images of Catphan500/503/504/600 were classified according to functional modules and studied by convolutional neural networks（CNN）. After training, the newly entered CT/CBCT images were automatically identified and sorted by functional modules, and then the related indicators were automatically analyzed, including HU linearity, modulation transfer function and uniformity of those images, aiming to ensure that the image quality met the requirements of clinical application.Results For the CT images of Catphan500 and the CBCT images of Catphan503, the function modules, including CTP401,CTP404 and CTP528, were correctly marked by CNN automatic classification. However, the accuracy of CTP486 didn＇t reach100%, which indicated that some other modules were wrongly classified into CTP486. Meanwhile, the automatic analysis of HU linearity, modulation transfer function and homogeneity of CT was achieved. Conclusion Based on CNN, CT/CBCT images of Catphan can be classified accurately. The next step will be to extend the method to other imaging devices in order to achieve a wider range of automatic image quality assurance.