Evaluation of Machine Learning methods for image denoising for dosimetry in Molecular Radiotherapy

1) Brief description of the context of the research including potential impact

[177Lu]Lutetium-rhPSMA-10.1 is a promising radioligand suitable for treating patients diagnosed with metastatic, castration-resistant prostate cancers. The therapeutic beta--particles (maximum energy 498 keV) are suitable for molecular radiotherapy due to their short range in tissue of 2-10 mm [Blue Earth Therapeutics, 2023]. Two gamma-emissions with energies of 113 keV (6.6% abundance) and 208 keV (11% abundance) enable single photon emission tomography (SPECT)/CT imaging to visualise and quantify uptake of the radiopharmaceutical [Dash, Pillai and Knapp Jr., 2015].

Post-therapy imaging is a crucial part of the theranostic pathway. SPECT/CT imaging is a regulatory requirement to enable quantification of radiopharmaceutical uptake within tumour volumes and organs at risk, and therefore estimate delivered radiation dose [EU directive 2013/59/Euratom]. SPECT/CT imaging of therapeutic radionuclides can be challenging, and data may be technically poor.

The aim of this project is to implement a machine learning algorithm which can be trained to denoise low-count reconstructed 177Lu-PSMA images. The goal is to generate images which more accurately represent the radiopharmaceutical distribution, both visually and quantitatively. A successful model could potentially be trained to de-noise clinical SPECT studies to increase quantitative accuracy and enable estimation of doses delivered to tumour volumes and organs at risk.

2) Aims and Objectives

- Complete a literature review to evaluate previous work, including potential deep learning model architectures which might be suitable for the project.
- Create three-dimensional digital 'phantoms' which mimic high- and low-count 177Lu-PSMA distributions which could be encountered in the clinic.
- Simulate and reconstruct phantom acquisitions using clinically realistic parameters, to generate a dataset of 'noisy' and 'noise-free' training samples.
- Select and train a machine learning model to produce 'noise-free' reconstructed images, using 'noisy' reconstructed images as input. Evaluate the model performance using traditional image quality metrics, including those relevant to SPECT quantification.
- Establish a starting point for the PhD project "Using AI to enable imaging of exotic radionuclides for Molecular Radiotherapy." The focus of the PhD will be to develop novel image reconstruction methods to enable quantitative post-therapy imaging for patients who have received molecular radiotherapy with a different radiopharmaceutical, [225Ac]Actinium-rhPSMA.

3) Novelty of Research Methodology

Radiotheranostics is a rapidly developing field in Nuclear Medicine. Artificial Intelligence offers the potential to extract quantitative data from low-count studies. This could improve patient-specific dosimetry accuracy and lead to a better understanding of the relationship between molecular radiotherapy dose and treatment outcomes.

4) Alignment to EPSRC's strategies and research areas

Radiotheranostics and molecular radiotherapy dosimetry are considered necessary for progression towards personalised medicine. This project involves using artificial intelligence to improve the accuracy and efficiency of medical image analysis for the purpose of diagnosis, treatment planning and monitoring.


5) Any companies or collaborators involved

Blue Earth Therapeutics (BET)
National Physics Laboratory (NPL)
University College London Hospitals NHS Foundation Trust (UCLH)