【作者】 李明欣；

【导师】 蒲立新；

【作者基本信息】 电子科技大学 ， 控制科学与工程， 2024， 硕士

【摘要】 计算机断层扫描(Computed Tomography,CT)技术能够展现病人体内细致的三维结构信息,为医生提供了更精确的诊断依据。其显著优势在于能够在三维空间中清晰呈现各类组织器官,有效避免了信息混叠的问题。然而,相较于X光检查,CT检查需要患者承受更高剂量的辐射,对身体健康构成较大威胁。鉴于此,本文利用深度学习技术,仅需两张正交的胸部X光片,便能重建出完整的胸部CT影像。为实现这一目标,我们设计了一种高效的融合模块,能够将二维的X光片数据提升至三维CT层面,并充分融合双视角的信息。同时,在特征从浅层向深层传递的过程中,该融合模块能够有效解决浅层与深层特征融合的问题。针对没有CT扫描设备的地区,本文提出了一种经济实惠的替代方案,只需利用低成本的DR采集设备,即可获得类似CT图像。实验证明,这种方法具有明显的有效性。本文的主要工作和贡献有:(1)在神经网络中,浅层的二维特征扮演着至关重要的角色。它们往往包含了图像中更为具体和细致的边缘轮廓、位置以及明暗等信息。这些信息对于后续的图像重建过程具有显著的促进作用,因为它们提供了重建目标的基础结构和细节线索。本文提出一种用于融合特征的编解码网络并命名为Xray2CT-CNN,模型巧妙结合CGAFusion模块与SKFusion模块,最后通过改进三维投影损失,提高模型在细节部分的重建能力。(2)由于Xray2CT-CNN的CT重建影像的质量仍有提升空间,本文将其作为生成对抗网络的生成器重建高质量CT影像。进而将扩散模型引入CT重建领域,结合扩散模型与生成对抗网络的优点,提出了一种高效的适用于端到端实现双视角胸部X光片重建CT影像的模型,命名为Xray2CT-DFGAN。模型可以充分利用两个视角的互补信息,生成高质量CT重建影像。(3)由于目前成对的X光片与CT公开数据量依然不足以训练出能力强大的模型。为此,本文在LIDC-IDRI数据集的基础上,通过在各个医院采集Hospital＿CT数据集,其中数据均已脱敏。借助于Cy Tran模型完成增强CT与平扫CT的风格转换。对缺少X光片的CT数据使用平均密度投影方法(Average Intensity Projection,AIP)技术生成虚拟的X光片。其次,对虚拟X光片进行对比度的增强,令图像风格尽可能地贴近真实X光片。最后对子数据集进行标准化。更多还原

【Abstract】 Computed Tomography(CT)technology displays detailed three-dimensional structural information in the patient’s body,giving doctors a more accurate diagnosis basis.Its significant advantage is that it can clearly present various tissues and organs in three-dimensional space,effectively avoiding the problem of information aliasing.However,compared with X-ray examination,CT examination requires patients to undergo higher doses of radiation,which poses a greater threat to health.In view of this,this thesis uses deep learning technology to reconstruct a complete chest CT image with only two orthogonal chest X-rays.To achieve this goal,we designed an efficient fusion module that can elevate the two-dimensional X-ray data to the three-dimensional CT level and fully integrate the dual-view information.At the same time,the fusion module can effectively solve the problem of shallow and deep feature fusion in the process of feature transmission from shallow layer to deep layer.For areas where CT scanning equipment is not available,this thesis proposes an economical alternative that can obtain CT-like images by using only low-cost DR Acquisition equipment.The experiment shows that this method has obvious effectiveness.The main work and contributions of this thesis are as follows:(1)In neural networks,shallow two-dimensional features play a crucial role.They often contain the more specific and detailed edge outline,position,and light and dark information in the image.This information has a significant role in the subsequent image reconstruction process,because it provides the basic structure and details of the reconstruction target.In this thesis,a feature fusion codec network named Xray2 CTCNN is proposed.The model skillfully combines CGAFusion module and SKFusion module,and finally improves the reconstruction ability of the model in detail by improving the 3D projection loss.(2)Since the quality of CT reconstructed images of Xray2CT-CNN still has room for improvement,this thesis uses Xray2CT-CNN as the generator of generative adversative network to reconstruct high-quality CT images.Then,the diffusion model was introduced into the field of CT reconstruction.Combining the advantages of the diffusion model and the generative adversal network,an efficient dual-view chest X-ray reconstruction CT image model named Xray2CT-DFGAN was proposed.The model can make full use of the complementary information of the two perspectives to generate high-quality CT reconstruction images.(3)Due to the current amount of paired X-ray and CT public data is still not enough to train a powerful model.Therefore,based on the LIDC-IDRI dataset,this thesis collects Hospital＿CT dataset from various hospitals,in which all the data have been desensitized.The style conversion of enhanced CT and plain CT was completed by Cy Tran model.The Average Intensity Projection(AIP)technique was used to generate virtual X-ray images of CT data without X-ray images.Secondly,the contrast of the virtual X-ray is enhanced to make the image style as close as possible to the real X-ray.Finally,the sub-data sets are standardized.更多还原