基于自注意力神经辐射场的稀疏视角医学影像重建方法  

A Method for Sparse-view Medical Image Reconstruction Based on Self-attention Neural Radiance Fields

作　　者：廖浩霖 李斯 Liao Haolin;Li Si(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006

出　　处：《广东工业大学学报》2025年第1期60-69,共10页Journal of Guangdong University of Technology

基　　金：国家自然科学基金资助项目(11771464);广东省自然科学基金资助项目(2022A1515012379)。

摘　　要：稀疏视角断层重建对于临床实践中减少辐射剂量具有重要意义。近年来,隐式神经表示(Implicit Neural Representation,INR)方法在医学图像的稀疏视角断层重建领域得到广泛应用,并展示了卓越的性能。然而,传统INR方法将每个采样点单独作为输入,忽视了邻近采样点之间的内在联系,从而削弱了重建效果。为解决上述问题,本文提出了一种新的INR方法。该方法首先将相邻射线上的邻近采样点重新组织成一系列感兴趣窗口,随后将上述感兴趣窗口输入到一个配备跳跃连接的Transformer查询网络中。利用Transformer网络的自注意力机制,该方法能够深入学习各个感兴趣窗口内采样点之间的内在联系,从而显著提升重建图像的质量。本文在锥束计算机断层扫描(Cone-beam Computed Tomography,CBCT)和平行束单光子发射计算机断层扫描(Single-photon Emission Computed Tomography,SPECT)两种断层成像模态上进行了大量的数值实验,与现有的先进INR方法Freq-NAF相比,本文所提方法在稀疏视角条件下的重建精度和图像可视化方面均表现出更优异的性能。特别是在胸部CBCT数据集,所提方法的峰值信噪比(Peak Signal-to-noise Ratio,PSNR)比Freq-NAF方法提高了0.45 dB。Sparse-view tomographic reconstruction is of significant importance for reducing radiation dose in clinical practice.In recent years,Implicit Neural Representation(INR)methods have been widely applied to medical image reconstruction in sparse-view scenario and have achieved competitive performance.However,traditional INR methods treat each sampling point individually as input,which neglect the inherent relations among neighboring sampling points,thus weakening the reconstruction performance.To address this,this paper proposes a novel INR method.The proposed method reorganizes neighboring sampling points on adjacent rays into multiple windows-of-interest,which are then fed into a Transformer query network equipped with a skip connection.By leveraging the self-attention mechanism of the Transformer network,the proposed method is able to capture the intrinsic relations among sampling points within each window-of-interest,thereby effectively enhancing the reconstructed image quality.This paper conducts extensive numerical experiments in two tomographic imaging modalities:Cone-Beam Computed Tomography(CBCT)and parallel-beam Single-Photon Emission Computed Tomography(SPECT).The experimental results show that,compared to the advanced INR method Freq-NAF,the proposed method achieves superior performance in terms of reconstruction accuracy and image visualization under sparse-view conditions,particularly obtaining a 0.45 dB improvement in Peak Signal-to-Noise Ratio(PSNR)on the chest CBCT dataset.

关 键 词：隐式神经表示 稀疏视角 医学影像重建 自注意力机制 感兴趣窗口 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]