da Costa-Luis Casper O, Reader Andrew J
Department of Biomedical EngineeringSchool of Biomedical Engineering and Imaging Sciences, St. Thomas' HospitalKing's College LondonLondonSE1 7EHU.K.
IEEE Trans Radiat Plasma Med Sci. 2020 Apr 8;5(2):202-212. doi: 10.1109/TRPMS.2020.2986414. eCollection 2021 Mar.
Noise suppression is particularly important in low count positron emission tomography (PET) imaging. Post-smoothing (PS) and regularization methods which aim to reduce noise also tend to reduce resolution and introduce bias. Alternatively, anatomical information from another modality such as magnetic resonance (MR) imaging can be used to improve image quality. Convolutional neural networks (CNNs) are particularly well suited to such joint image processing, but usually require large amounts of training data and have mostly been applied outside the field of medical imaging or focus on classification and segmentation, leaving PET image quality improvement relatively understudied. This article proposes the use of a relatively low-complexity CNN (micro-net) as a post-reconstruction MR-guided image processing step to reduce noise and reconstruction artefacts while also improving resolution in low count PET scans. The CNN is designed to be fully 3-D, robust to very limited amounts of training data, and to accept multiple inputs (including competitive denoising methods). Application of the proposed CNN on simulated low (30 M) count data (trained to produce standard (300 M) count reconstructions) results in a 36% lower normalized root mean squared error (NRMSE, calculated over ten realizations against the ground truth) compared to maximum-likelihood expectation maximization (MLEM) used in clinical practice. In contrast, a decrease of only 25% in NRMSE is obtained when an optimized (using knowledge of the ground truth) PS is performed. A 26% NRMSE decrease is obtained with both RM and optimized PS. Similar improvement is also observed for low count real patient datasets. Overfitting to training data is demonstrated to occur as the network size is increased. In an extreme case, a U-net (which produces better predictions for training data) is shown to completely fail on test data due to overfitting to this case of very limited training data. Meanwhile, the resultant images from the proposed CNN (which has low training data requirements) have lower noise, reduced ringing, and partial volume effects, as well as sharper edges and improved resolution compared to conventional MLEM.
噪声抑制在低计数正电子发射断层扫描（PET）成像中尤为重要。旨在降低噪声的后平滑（PS）和正则化方法往往也会降低分辨率并引入偏差。另外，可以使用来自另一种模态（如磁共振（MR）成像）的解剖学信息来提高图像质量。卷积神经网络（CNN）特别适合这种联合图像处理，但通常需要大量的训练数据，并且大多应用于医学成像领域之外，或者专注于分类和分割，相对而言对PET图像质量改善的研究较少。本文提出使用一种复杂度相对较低的CNN（微网络）作为重建后MR引导的图像处理步骤，以减少噪声和重建伪影，同时提高低计数PET扫描的分辨率。该CNN设计为全三维，对非常有限的训练数据具有鲁棒性，并能接受多个输入（包括竞争性去噪方法）。将所提出的CNN应用于模拟的低（30 M）计数数据（训练以生成标准（300 M）计数重建），与临床实践中使用的最大似然期望最大化（MLEM）相比，归一化均方根误差（NRMSE，针对十个实现相对于真实情况计算）降低了36%。相比之下，当执行优化的（利用真实情况的知识）PS时，NRMSE仅降低25%。使用RM和优化的PS时，NRMSE降低了26%。对于低计数真实患者数据集也观察到了类似的改善。随着网络规模的增加，证明会出现对训练数据的过拟合。在一个极端情况下，一个U网络（对训练数据产生更好的预测）由于对这种非常有限的训练数据情况的过拟合，在测试数据上完全失败。同时，与传统的MLEM相比，所提出的CNN（对训练数据要求较低）生成的图像具有更低的噪声、减少的振铃和部分容积效应，以及更清晰的边缘和更高的分辨率。
