Convolutional neural networks (CNNs) have recently achieved state-of-the-art results for positron emission tomography (PET) imaging problems. However direct learning from input image to target image is challenging if the gap is large between two images. Previous studies have shown that CNN can reduce image noise, but it can also degrade contrast recovery for small lesions. In this work, a deep progressive learning (DPL) method for PET image reconstruction is proposed to reduce background noise and improve image contrast. DPL bridges the gap between low quality image and high quality image through two learning steps. In the iterative reconstruction process, two pre-trained neural networks are introduced to control the image noise and contrast in turn. The feedback structure is adopted in the network design, which greatly reduces the parameters. The training data come from uEXPLORER, the world’s first total-body PET scanner, in which the PET images show high contrast and very low image noise. We conducted extensive phantom and patient studies to test the algorithm for PET image quality improvement. The experimental results show that DPL is promising for reducing noise and improving contrast of PET images. Moreover, the proposed method has sufficient versatility to solve various imaging and image processing problems.

卷积神经网络 (CNN) 最近在正电子发射断层扫描 (PET) 成像问题上取得了最先进的结果。然而，如果两幅图像之间的差距很大，从输入图像到目标图像的直接学习是具有挑战性的。之前的研究表明，CNN 可以减少图像噪声，但它也会降低小病变的对比度恢复。在这项工作中，提出了一种用于 PET 图像重建的深度渐进学习 (DPL) 方法，以减少背景噪声并提高图像对比度。DPL 通过两个学习步骤弥合了低质量图像和高质量图像之间的差距。在迭代重建过程中，引入两个预训练的神经网络依次控制图像噪声和对比度。网络设计采用反馈结构，大大减少了参数。训练数据来自uEXPLORER，世界上第一台全身PET扫描仪，其中PET图像显示出高对比度和极低的图像噪声。我们进行了广泛的体模和患者研究，以测试算法是否能提高 PET 图像质量。实验结果表明，DPL 有希望降低噪声和提高 PET 图像的对比度。此外，所提出的方法具有足够的通用性来解决各种成像和图像处理问题。