正电子发射断层成像（Positron Emission Tomography, PET）技术是目前在临床上应用广泛的一种核医学分子影像技术，主要用于对肿瘤的筛查与诊断。PET扫描过程中会给患者注射放射性药物来进行成像，为了减少患者受到的辐射剂量，在注射低剂量放射性药物的条件下进行高质量的PET成像具有重要的研究意义和临床应用价值。低剂量放射性药物成像带来的首要问题是图像噪声增大。本文利用PET低剂量与高剂量图像构成的数据集对U-Net神经网络进行训练，以得到能够有效对低剂量图像降噪的网络模型，并通过修改病灶及邻域附近的损失函数来增加对病灶的训练强度，经过对模型预测图像的信噪比、相对偏差、归一化均方误差等质量评估指标的计算，认为构建的模型能够对低剂量图像进行有效降噪，在提高信噪比的同时能保证对病灶的定量精度，保证图像信息精确度与完整度。本工作对在低剂量条件下的PET成像研究具有一定的参考价值。

Positron emission tomography (PET) is a widely used nuclear medicine molecular imaging technique in clinical practice, mainly for screening and diagnosis of tumors. During PET scanning, radioactive drugs are injected into patients for imaging. To reduce the radiation dose that patients receive, it is of great research significance and clinical application value to perform high-quality PET imaging under the condition of injecting low-dose radioactive drugs. The primary problem caused by low-dose radioactive drugs is the increase of image noise. This paper uses a dataset composed of PET low-dose and high-dose images to train a U-Net, to obtain a network model that can effectively denoise low-dose images, and modifies the loss function near the lesion and its adjacent areas to increase the training intensity for the lesion. After calculating the quality evaluation indicators such as signal-to-noise ratio, relative deviation, and normalized mean square error of the model predicted image, this paper believes that the constructed model can effectively denoise the low-dose images, and ensure the quantitative accuracy of the lesions and the accuracy and completeness of the image information while improving the signal-to-noise ratio. This work has certain reference value for PET imaging research under low-dose conditions.