Automated Assessment of Clinical Brain PET Image Quality

Abstract

It is often desirable to reduce the quantity of radioligand used in Positron Emission Tomography (PET), in order to reduce the radiation exposure to both patients and staff. Using a fraction of the standard amount of the radioligand reduces radiotracer costs. This, however, results in a fraction of the usual photon counts being available, with the reconstructed image quality becoming relatively poor due to increased noise, which in turn often leads to a loss of resolution. To ensure that the reconstructions can be used in a clinical setting, an image quality assessment is therefore essential, especially from a clinical perspective, rather than (or in addition to) conventional image quality metrics.

This thesis proposes and evaluates automated image quality assessment Convolutional Neural Networks (CNNs) for [¹⁸F]FDG brain PET images using clinically-designed metrics: global quality rating, pattern recognition, and diagnostic confidence. Real data collected from memory clinic patients with suspected dementia were reconstructed at varying simulated injected dose levels (0.5% - 100% of counts). These automated image quality assessment tools aim to aid large-scale reconstruction investigations, with a CNN used as an independent observer. The two main application areas considered specifically in this thesis are assessing clinical viability of low-count standard image reconstructions, and novel reconstruction methods suited to low-count data.

One of the primary issues with the use of deep learning in medical imaging is the relative paucity of clinically-annotated medical images, arising from limited time availability of clinicians to read the images. This thesis proposes utilising CNNs pre-trained on natural image data with transfer learning to brain [¹⁸F]FDG PET images, to mitigate the impact of limited clinically-annotated data. A specific pre-trained CNN, VGG16, was initially investigated as both a function of the amount of training data used, and of the number of parameters that were fine-tuned.

This method was then applied to other pre-trained CNN backbones, to better understand the driving forces behind model improvement, such that this knowledge can be used for other studies. It was found that training between ~6,000 – 20,000 parameters in combination with an “off-the-shelf” pre-trained CNN model provided the lowest mean-absolute-error. It was also found that unfreezing the parameters of the last 40% of layers improved model performance, suggesting that an over-parameterised regime may work better for this clinical task, compared to models with a more conventional level of parameterisation.

When the trained models were applied to novel reconstruction methods, an MR-guided regularised reconstruction method with post-smoothing outperformed standard Maximum Likelihood Expectation Maximisation (MLEM) reconstruction at reduced count levels (5% of the standard injected dose), in accordance with human observers. Overall, the CNN models yielded similar trends to the clinical readings, and as such have potential to be used as independent observers. This may aid, for example, in assessing clinical feasibility for low-count standard reconstructions, or for assessing novel reconstruction methodology, thereby assisting in large-scale clinical image quality assessment investigations.

Date of Award	1 Jan 2025
Original language	English
Awarding Institution	King's College London
Supervisor	Alexander Hammers (Supervisor) & Andrew Reader (Supervisor)