【作者】 吴海涛；

【导师】 刘爽；

【作者基本信息】 河北大学 ， 仪器科学与技术， 2022， 硕士

【摘要】 指掌骨骨折是一种常见的突发性骨疾病,一般需要通过急诊下X光成像进行快速定位和治疗,若处理不及时或治疗不恰当,可能会导致患者恢复困难或者是预后不佳等情况。然而,指掌骨骨折的部位特殊,关节不仅多而且很小,骨折位置很隐匿,不容易被识别,容易造成误诊,漏诊的情况。随着数据聚合和人工智能技术的快速发展,基于深度学习的卷积神经网络模型被广泛应用在很多领域,尤其是在医学影像检测领域,并且取得了很好的效果。因此使用深度学习模型来实现对指掌骨部位微小骨折的检测并积极探索应用于临床,具有重要的研究意义和应用价值。本文研究了一种改进的Faster RCNN模型对指掌骨微小骨折进行检测。主要研究内容如下:第一,指掌骨X光图像的收集,预处理,扩充,标注,扩充后的整体数据量达到了5195张,满足了深度学习算法对数据量的要求。同时对原始数据集中由仪器性能不高或者是拍摄角度不合适造成的低分辨率模糊图像进行预处理。进行实验对比后的结果表明,基于双三次方插值的超分辨率算法处理之后的图像的平均峰值信噪比能达到34.8871 d B,平均结构相似性能够达到0.9647,优于其它算法的处理效果。最终选定了此方法实现对低分辨率图像的预处理,提升图像的分辨率,同时也能保证处理后的图像和原始图像在结构上没有很大的偏差,为后续骨折的检测实验提供了高质量的数据集。第二,根据数据集进行了网络模型训练策略的研究。通过训练次数和参数调节的实验,找到适合的实验训练次数21次和非极大值抑制的值0.7,同时对原始的Faster RCNN模型的Res Net50主干网络,FPN网络,RPN网络和ROI网络的参数进行设置,配置后的模型在预处理后的数据集上的平均精确度能够达到65.5%,为后续改进模型的参数配置提供参考。第三,针对Res Net50在输入阶段只能提取图像单一尺度的特征,导致特征图中多尺度信息不够丰富的问题,在改进的主干网络结构ISSC-Res Net50中设计了双输入路径的input-stem-short-cut结构。该结构在Res Net50输入主干中加入了一条新路径,由1*1卷积层和4*4最大池化层组成,可以用不同的卷积采样方式来处理输入图像。两条路径同时工作,并将输出结果进行融合。因此,与常规输入主干相比,input-stem-short-cut的结构可以获取输入图像的不同尺度的丰富信息。在增加input-stem-short-cut结构的Faster R-CNN模型中的平均精确度达到了66.4%,比原始模型的平均精确度提高了0.9%。第四,原始的下采样模块中的shortcut路径中采用的是1*1的卷积,因此忽略了输入特征图四分之三的信息,导致特征图的提取效果受到了很大的影响。本文对此问题进行了解决,采用了2*2的卷积降低了输入特征图的信息损失,有助于对输入图像信息的全面采集。由于这种下采样模块在整体的ISSC-Res Net50的结构中数量很少,因此在提高网络检测性能的基础上几乎不增加参数量,并且也不会影响整个网络的速度。在改变下采样模块的Faster R-CNN模型中的平均精确度也达到了66.4%,比原始模型提高0.9%。同时针对原始模型中NMS和ROIPooling检测精度不高的问题进行了解决。将其替换为Soft NMS和ROIAlign,模型的平均精确度提高到了66.1%。根据实验表明,整体改进模型在预处理后的数据集上的平均精确度能够达到67.7%,同时模型损失比原始模型降低,实现了对指掌骨部位微小骨折检测性能的提升。通过对Faster R-CNN的改进,在其基础上获得了性能和检测效果更好的指掌骨骨折检测模型,根据实验证明了对指掌骨骨折的检测精确度得到了提升,同时本文改进的模型和其它经典的模型相比,都取得了很好的效果,因此对于实际的临床应用具有一定参考意义。更多还原

【Abstract】 Phalangeal and metacarpal fracture is a common sudden bone disease,which generally needs to be quickly located and treated by emergency X-ray imaging.If the treatment is not timely or inappropriate,it may lead to difficult recovery or poor prognosis.However,the location of phalangeal and metacarpal fracture is special,the joints are not only many but also very small,and the fracture location is hidden,which is not easy to be identified.It is easy to cause misdiagnosis and missed diagnosis.With the rapid development of data aggregation and artificial intelligence technology,convolution neural network models based on deep learning are widely used in many fields,especially in the field of medical image detection,and have achieved good results.Therefore,using the deep learning model to detect the micro fracture of phalangeal and metacarpal bone and actively explore its application in clinic has important research significance and application value.In this paper,an improved Faster R-CNN model is studied to detect the micro fracture of phalangeal and metacarpal bone.The main research contents are as follows:First,the collection,preprocessing,expansion and labeling of phalangeal and metacarpal X-ray images.The number of expanded data sets has reached 5195,which meets the requirements of deep learning algorithm.At the same time,the low-resolution blurred images caused by the low performance of the instrument or the improper shooting angle in the original data set are preprocessed.After experimental comparison,the results show that the average peak signal-to-noise ratio of the image processed by the super-resolution algorithm based on bicubic interpolation can reach 34.8871 d B,and the average structural similarity can reach0.9647,which is better than other algorithms.Finally,this method is selected to preprocess the low-resolution image and improve the resolution of the image.At the same time,it can also ensure that there is no great deviation in the structure between the processed image and the original image,which provides a high-quality data set for the subsequent fracture detection experiment.Secondly,the training strategy of network model is studied according to the data set.Through the experiments of training times and parameter adjustment,the suitable experimental training times of 21 times and the value of non maximum suppression of 0.7 are found.At the same time,the parameters of Res Net50 backbone network,FPN network,RPN network and ROI network of the original Faster R-CNN model are set.The average precision of the configured network on the preprocessed data set can reach 65.5%,which provides a reference for the subsequent improvement of model parameter configuration.Thirdly,aiming at the problem that Res Net50 can only extract the features of a single scale of the image in the input stage,resulting in insufficient multi-scale information in the feature map,a double input path input-stem-short-cut structure is designed in the improved backbone network structure ISSC-Res Net50.This structure adds a new path to the Res Net50 input stem,which is composed of 1 * 1 convolution layer and 4 * 4 maxpooling layer.Different convolution sampling methods can be used to process the input image.The two paths work at the same time,and the output results are fused.Therefore,compared with the conventional input stem,the input-stem-short-cut structure can obtain rich information of different scales of the input image.The average precision of Faster R-CNN model with input-stem-short-cut structure is 66.4%,which is 0.9% higher than that of the original model.Fourth,the 1 * 1 convolution is used in the shortcut path in the original downsampling block,so three quarters of the information of the input feature map is ignored,resulting in a great impact on the extraction effect of the feature map.This paper solves this problem by using2 * 2 convolution,which reduces the information loss of the input feature map and is helpful to the comprehensive acquisition of the input image information.Because this downsampling block has a small number in the overall structure of ISSC-Res Net50,it hardly increases the amount of parameters on the basis of improving the network detection performance,and will not affect the speed of the whole network.With the change of downsampling block,the average precision of Faster R-CNN model also reaches 66.4%,which is 0.9% higher than the original model.At the same time,the problem of low detection precision of NMS and ROIPooling in the original model is solved.Replacing them with Soft NMS and ROIAlign,the average precision of the model is improved to 66.1%.According to the experimental results,the average precision of the overall improved model on the preprocessed data set can reach 67.7%.At the same time,the loss of the model is lower than that of the original model,which improves the detection performance of micro fractures in the phalangeal and metacarpal bone.Through the improvement of Faster R-CNN,a phalangeal and metacarpal fracture detection model with better performance and detection effect is obtained.According to the experiment,it is proved that the detection precision of phalangeal and metacarpal fracture has been improved.At the same time,the improved model in this paper has achieved good results compared with several classical models.Therefore,it has a certain reference significance for practical clinical application.更多还原