首页|期刊导航|分子影像学杂志|基于改进深度图像先验的PET图像部分容积校正方法

基于改进深度图像先验的PET图像部分容积校正方法OA

Partial volume correction in PET imaging using improved deep image prior

中文摘要英文摘要

目的 提出一种基于改进深度图像先验的无监督小样本PET图像部分容积校正方法.方法 在传统深度图像先验框架上进行改进,提出了PET图像部分容积校正新方法,使用群体信息进行预训练,采用反投影保真项结合去噪正则化替代传统损失函数,引入神经盲反卷积估计系统点扩散函数,通过注意力机制融合MRI解剖特征,进一步提高校正有效性.结果 在PET-SORTEO模拟数据集与ADNI临床数据集开展对比实验,本方法的PSNR、SSIM、CNR及CLIP、NIQE图像质量评价指标表现均优于对比方法.结论 所提基于改进深度图像先验的PET图像部分容积校正方法,有效提升了PET图像质量.

Objective To propose an unsupervised few-shot PET partial volume correction method based on an improved deep image prior.Methods Propose a novel PET partial volume correction method by enhancing the classical deep image prior framework.Specifically,population information is used for pre-training;the conventional loss function is replaced with a back-projection fidelity term combined with regularization by denoising;neural blind deconvolution is introduced to estimate the system point spread function;MRI anatomical features are fused via an attention mechanism to further improve correction performance.Results Validation experiments were carried out on PET-SORTEO simulation dataset and ADNI clinical dataset.The test results of quantitative indicators showed that the PSNR,SSIM,CNR,CLIP and NIQE of the proposed method achieved superior performance over the compared approaches.Conclusion The proposed method,built upon the modified deep image prior for PET partial volume correction,effectively enhances the image quality of PET images.

刘政;高园园;傅蓉

南方医科大学生物医学工程学院,广东 广州 510515南方医科大学生物医学工程学院,广东 广州 510515南方医科大学生物医学工程学院,广东 广州 510515

部分容积校正深度图像先验神经盲反卷积注意力网络

partial volume correctiondeep image priorneural blind deconvolutionattention network

《分子影像学杂志》 2026 (8)

981-989,9

国家自然科学基金(32101108)Supported by National Natural Science Foundation of China(32101108).

10.12122/j.issn.1674-4500.2026.08.01

评论