三维多尺度嵌套U结构CT影像肺结节检测

Detection of pulmonary nodules in three-dimensional multiscale nested U-structure computed tomography images

目的肺结节检测在低剂量肺部计算机断层扫描(computed tomography,CT)筛查肺癌中具有重要意义。但由于结节大小、形状和密度的变化十分复杂,导致难以在低假阳性率下保证高的灵敏度,这限制了深度学习算法在常规临床实践中的肺结节自动诊断,建立具有良好结节检测性能的深度学习模型仍然是一个挑战。针对此问题,本文提出了一种基于3D ReSidual U(3D RSU)块的嵌套U结构的肺结节检测框架。方法3D RSU块通过混合不同大小的感受场获得多尺度特征来丰富特征信息。而嵌套U结构允许网络获得更大分辨率的特征图,从而具有多层次深度特征,获得丰富的局部和全局信息,增强网络区分前景和背景的能力,进而提高微小结节等非显著性目标的检测性能。结果该框架在公共肺结节(lung nodule analysis 16)挑战数据集上进行了评价。方法能够准确地检测出肺结节,灵敏度达到了97.2%,与基准方法相比,该方法灵敏度提高了2.6%,具有很高的灵敏度和特异性,在0.125、0.25、0.5、1、2、4、8共7个假阳率点的灵敏度平均值为86.4%,尤其是在每扫描0.25个假阳性上,灵敏度达到80.9%,优于基准算法76.9%的结果。结论本文所提出的结节检测模型,由于在低假阳率上有较高的灵敏度,可以使本网络在常规临床实践中为医生辅助诊断提供更加可靠清晰的早期肺癌参考信息。

Objective Lung cancer is one of the most common diseases in humans and mainly causes the rising mortality rate.Medical experts believe that the early diagnosis of lung cancer can reduce mortality by screening for lung nodules through computed tomography(CT).Checking a large number of CT images can reduce lung cancer risk.However,CT scan images contain a large volume of information about nodules,and as the number of images increases,accuracy becomes a very challenging task for radiologists.Therefore,an effective computer-aided diagnosis(CAD)system should be designed.However,the high false-positive rate remains a challenging problem.In response,this paper proposes a new lung nodule detection framework based on 3 D ReSidualU-blocks(3 D RSU)module and nested U structure.Method This paper trains an end-to-end lung nodule detection model for one stage.The whole system can shorten the processing time of pulmonary nodule detection without reducing the accuracy of early detection.The 3 D deep convolutional neural network(CNN)proposed in this paper is based on the region proposal network(RPN)of the Faster R-CNN Network.The 3 D CNN can make full use of the spatial information of CT images.The problem of missed detection by the nodule and the large number of false-positive nodules during detection of nodules with small diameter nodules can be solved by dividing the front background information and enhancing the ability to detect non-salient objects in the image.In this paper,the 3 D RSU was designed to capture multi-scale features within the stage.The symmetrical codec structure can be used to learn how to extract and encode multi-scale context information.By using the different layers of 3 D RSU,the network can allow feature maps of any spatial resolution as input elements to extract elements at multiple scales.This process reduces the loss of detail caused by large-scale direct up-sampling.The 3 D RSU was embedded into the network to form a nested u-shaped structure.This structure allows the network to obtain larger resolution feature maps,thus providing multiple levels of deep features.High-level semantic features can be transitioned to supplement low-level conventional features.Improve the detection performance of non-important targets such as calcification and ground glass.If high-resolution feature maps only are used for prediction,a large amount of lung nodule position information contained in low-resolution feature maps will be lost.Therefore,a feature pyramid network was designed in the decoding structure to integrate low-level high-resolution features.In addition to the strong high-level semantic features,the location details and strong semantics of nodule detection are enriched,and the detection of small nodules is realized.Result The framework was evaluated on the public lung nodule analysis(LUNA16)challenge data set containing 888 patient lung nodule labels.A total of 10 folds were observed in the entire data set.This article used a 10-fold cross-validation method to compare performance indicators.Our network training uses a stochastic gradient descent optimizer.The initial learning rate is 0.01.After reaching 1/3 and 2/3 of the total number of iterations,the learning rate was adjusted to 0.001 and 0.0001,respectively,and the weight attenuation is 1×10;.The batch size is 8,the number of model iterations is 120,one round of training takes approximately 620 s,and the training of a model takes approximately 20 h.Through ablation and comparative experiments,this paper discusses the advantages of 3 D RSU module in pulmonary nodule detection performance compared with Res18,and the structural advantages of nested U structure over traditional U structure.Fewer false-positives means more correct nodules are identified with fewer errors,and more immediate help for doctors.Compared with the 3 D ResNet-18 module,the 3 D RSU module proposed in this paper has a significant improvement in low false positives.It shows that the 3 D RSU module reduces the detail loss caused by large-scale direct-up sampling through codec structure and cascade operation.The semantic features of pulmonary nodules were used to obtain the highest number of low-level features of pulmonary nodules,to allow the network to directly extract multiple proportion features from each residual block.Therefore,compared with 3 D Resnet-18 block,3 D RSU block can obtain feature map with richer semantics and clearer location information.Notably,a low false-positive rate is important0 for the system to identify acceptable modules with less false-positive results.Therefore,the 3 D RSU module has a more extensive clinical application value.In comparison with the traditional U-NET structure,the nested U structure proposed in this paper has a slightly higher parameter value than the Res18_FPN_RPN network,and the gap in the test time is almost negligible,indicating that although the structure of this article is more complex,the nested U structure does not remarkably.Increase the computational overhead.In comparison with benchmark experiments,our method can accurately detect lung nodules with a sensitivity of 97.1%,and the sensitivity increased by 2.5%.This method has high sensitivity and specificity.The average sensitivity of the seven false alarm rate points of 0.125,0.25,0.5,1,2 and 8 is 86.4%,especially when 0.25 false alarms are scanned per scan,in which the sensitivity reaches 80.9%,which is better than the result of 76.9%.Conclusion The nodule detection model proposed in this paper has high sensitivity in the low false-positive rate,thus allowing this network to provide more reliable and clear reference information for early lung cancer diagnosis in routine clinical practice for doctors.