Low-dose PET/MR image enhancement using deep learning with an MR-based anatomical structure attention mechanism

Positron emission tomography (PET) is a vital tool for molecular-level imaging but is limited by high radiation exposure, prompting the development of low-dose PET techniques. However, low-dose PET images often exhibit increased noise, which reduces their diagnostic utility. Magnetic resonance (MR) imaging, renowned for its superior soft tissue contrast and absence of ionizing radiation, offers potential for enhancing low-dose PET quality. Deep learning methods, particularly convolutional neural networks (CNNs) and transformers, have shown promise in improving image quality through multi-modal data integration, yet challenges remain in effectively leveraging MR for PET denoising.

To improve image quality of low-dose PET while minimizing radiotracer dosage by employing deep learning methods in conjunction with MRI.

In this retrospective study, PET/MR data from 59 patients were used for training and validation. Merge Net, which utilizes a Unet++ architecture enhanced with a self-attention mechanism and takes registered low-dose PET and MR images as input, was used to synthesize full-dose PET images. To evaluate the contribution of MR images to PET denoising, we conducted the following comparative experiments: the PET_MR model used paired low-dose PET and MR images as input, while the PET_only model relied solely on low-dose PET images. Both models shared the same network architecture to ensure a fair comparison. Image quality was assessed using peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and root mean square error (RMSE) relative to full-dose PET images. Two readers scored subjective image quality using a 5-point Likert scale (5 = excellent, 1 = very poor). Statistical analyses were carried out to assess differences in image quality metrics and the detection of lesions.

Compared with the low-dose images, the enhanced images, particularly those from the PET_MR model, demonstrated significant improvements across all quality metrics. Quantitative results from 12 test cases demonstrated high similarity between the enhanced low-dose PET images and the full-dose PET images (PSNR = 33.9, SSIM = 0.87, RMSE = 2.08 for the PET_MR model; PSNR = 33.43, SSIM = 0.86, RMSE = 2.18 for the PET_only model). Visually, the PET_MR model demonstrated clearer tissue boundaries and reduced noise levels compared with the PET_only model. All PET_MR images received scores of 3 or higher, comparable to those of the full-dose images.

The PET_MR model significantly enhanced image quality, offering clearer tissue boundaries and reduced noise.