Deep Learning-Driven Improvements in Low-Activity PET Imaging for Neurodegenerative Diseases

Introduction

This project explores the enhancement of PET imaging for diagnosing neurodegenerative disorders such as Alzheimer's and Parkinson's by reducing the amount of radioactive material (activity) or scanning time while maintaining diagnostic accuracy. By developing a Deep Learning model for denoising (noise reduction), PET images with standard-activity quality can be obtained from low-activity scans, aiming to minimize radiation exposure and optimize the patient experience.

Project Background

Neurodegenerative disorders, including Alzheimer's disease (AD) and Parkinson's disease (PD), primarily affect patients over the age of 60. Alzheimer's disease is characterized by amyloid deposition in the brain and can be imaged using various PET radiotracers, including [¹¹C]PiB, to confirm or rule out a clinical diagnosis. Similarly, specific PET tracers, such as [¹⁸F]FE-PE2I, can visualize the loss of dopaminergic neurons associated with Parkinson's disease, aiding in the diagnosis of patients.

However, current PET scanning techniques face obstacles such as high radiation exposure and long scanning times. Low-activity PET images often result in reduced image quality, creating a need for improvements in scanning technologies and methods. This project addresses these challenges by utilizing a Deep Learning model for noise reduction, aimed at improving the image quality of low-dose PET scans. This enables reliable diagnostics with reduced radiation exposure and shorter scanning times.

Project Potential

The employed model has demonstrated its ability to effectively reduce noise in simulated low-activity PET images, resulting in images that closely match those obtained with full activity. This represents a significant improvement, enabling the potential to reduce tracer dose or scanning time in PET/CT without compromising image quality. This indicates the potential for implementing these methods in clinical practice, which may facilitate more effective strategies for diagnosing and treating neurodegenerative disorders.

Contact Information

Name: Raphaël Sura Daveau & Claes Nøhr Ladefoged

Email: claes.noehr.ladefoged@regionh.dk

Location: Department of Clinical Physiology and Nuclear Medicine, Rigshospitalet, University of Copenhagen, Denmark

Position: ?

Publications

Deep learning based low-activity PET reconstruction of [11C]PiB and [18F]FE-PE2I in neurodegenerative disorders

The Deep Learning Model for PET Image Denoising

The Deep Learning model is engineered to generate standard-activity PET images from low-activity PET scans. It utilizes a three-dimensional U-Net architecture, which consists of three primary components: an encoder, a bridge, and a decoder, working in unison to facilitate the denoising process.

Figure 1: Example of PET Imaging — **Figur 1** illustrates the U-Net architecture employed to remove noise from low-activity PET images. The model takes two inputs: the low-activity PET image and a noise map that highlights the noise present in the image.
The inputs are processed through several convolutional blocks, with average pooling applied after each block to manage image resolution (Avg pool) while leveraging multiple filters for feature extraction. Following processing at the lowest resolution (the bridge), the image undergoes upsampling (Trans Conv) through transposed convolution to restore clarity.
To retain crucial information from earlier processing stages, features from the encoder are transferred to the decoder. Upon completion of the process, the model produces a new, denoised PET image, characterized by improved quality and reduced noise, derived from the original low-activity input. Improvements to low-activity PET images with the noise-reduced model are shown in Figure 2.

Figure 2: Results of Low-Activity Scanning — **Figur 2:** PET images of the brain in the transverse plane using [¹¹C]PiB (associated with AD) and [¹⁸F]FE-PE2I (associated with PD). The images show examples of normal, borderline, and abnormal scans. Each panel displays standard-activity images (left), low-activity images (center), and denoised images (right). White markings highlight areas of potential amyloid accumulation for [¹¹C]PiB, relevant to Alzheimer's disease, and for [¹⁸F]FE-PE2I, where the putamen and caudate nucleus, associated with Parkinson’s disease, are highlighted. The model effectively cleanses the images of noise.