Clinical and Deep-Learned Evaluation of MR-Guided Self-Supervised PET Reconstruction

Jessica B. Hopson, Sam Ellis, Anthime Flaus, Colm J. McGinnity, Radhouène Neji, Andrew J. Reader, Alexander HammersIEEE Transactions on Radiation and Plasma Medical Sciences 2024 · DOI

TL;DR

This paper investigates novel self-supervised PET reconstruction methods, specifically MR-guided MAPEM, for low-dose brain imaging. They found that MR-guided MAPEM, even with a drastically reduced radiation dose (5% of standard), can achieve image quality comparable to standard MLEM reconstruction at full dose, making low-dose PET more clinically viable. Furthermore, they demonstrate that Convolutional Neural Networks (CNNs) can act as automated observers, mirroring clinical assessments and potentially speeding up large-scale evaluations.

What Problem Does This Paper Address?

Positron Emission Tomography (PET) is a crucial tool for functional brain imaging, especially in memory clinics. However, the use of radioactive tracers necessitates a radiation dose to patients, which has associated risks and costs. To mitigate this, researchers are exploring low-dose PET acquisitions. The challenge is that lower radiation doses lead to fewer detected events, resulting in degraded image quality with traditional reconstruction methods like MLEM. This degradation can hinder accurate diagnosis. While advanced reconstruction techniques exist, their clinical utility, especially at very low doses, needs rigorous evaluation from a clinical perspective. This paper aims to bridge this gap by clinically assessing MR-guided self-supervised PET reconstruction methods at low doses and comparing them to the current standard.

Methods

The study compares four PET reconstruction algorithms:

  • MLEM (Maximum-Likelihood Expectation-Maximization): The clinical standard, used as a benchmark. Reconstructions were performed at both 100% and 5% of the standard injected dose.
  • Self-supervised Unguided MAPEM: A self-supervised maximum a posteriori expectation-maximization algorithm that uses data bootstrapping for regularization.
  • Self-supervised MR-guided MAPEM: Similar to the unguided version, but it uses the patient's MR image to guide the regularization process, controlling the smoothness penalty. This is a key novelty.
  • Post-smoothed versions: For both MAPEM methods, smoothed versions were also evaluated, using a Gaussian filter matching the point-spread-function (PSF) used in resolution modeling.

The clinical evaluation involved two experienced clinicians assessing images based on seven metrics: global quality rating (GQR), pattern recognition (PR), diagnostic confidence (DC), sharpness, caudate-putamen separation (CP), noise, and contrast. Additionally, pre-trained Convolutional Neural Networks (CNNs) were employed to automate image quality assessment on a separate set of images.

Datasets

The study utilized PET-MR brain imaging data from 11 memory clinic patients (6 female, 5 male, average age ~58.5 years) who were part of a larger study. These patients underwent both PET-CT and PET-MR scans. For this study, only the PET-MR scans were used. The data was acquired with [<sup>18</sup>F]FDG as the radioligand. The reconstructions were performed using 100% and 5% of the total counts, with the 5% data obtained by resampling. A separate set of 62 images, read by a single clinician, was used for the CNN evaluation.

Key Insights and Ideas

  • MR Guidance is Key: The MR-guided MAPEM reconstruction significantly outperformed MLEM, especially for technical metrics like sharpness, caudate-putamen separation, and contrast, even at very low doses. This highlights the power of incorporating anatomical information from MRI into PET reconstruction.
  • Self-Supervision Reduces Reliance on External Data: The self-supervised nature of the MAPEM methods means they don't require large external datasets for training, making them more adaptable and potentially easier to implement clinically.
  • CNNs as Automated Observers: The study demonstrates that CNNs can effectively predict clinical image quality scores, mirroring human observer trends. This opens the door for automated, rapid assessment of reconstruction methods, crucial for large-scale investigations.
  • Clinical Viability at Low Doses: The findings suggest that with advanced reconstruction, PET imaging can be performed with significantly reduced radiation doses without a substantial loss in clinically relevant image quality.

Key Results

  • Clinical Evaluation: MR-guided MAPEM with post-reconstruction smoothing achieved statistically significant improvements over MLEM in sharpness, caudate-putamen separation, and contrast, even at 5% of the standard injected dose. For example, the score for caudate-putamen separation was 0.07 ± 0.12 for MLEM at 5% counts versus 0.92 ± 0.52 for MR-guided MAPEM.
  • Overall Performance: At 5% of the standard injected dose, MR-guided MAPEM with post-reconstruction smoothing was rated as "poor but useable" or "clinical standard" for 6 out of 7 metrics, compared to only 3 out of 7 for MLEM.
  • CNN Evaluation: CNN predictions showed similar trends to human observers, with MR-guided MAPEM with post-reconstruction smoothing performing best at 5% counts, matching MLEM at 100% counts.
  • Comparison: For 5% count data, MR-guided MAPEM with post-reconstruction smoothing was rated significantly better than MLEM for 3 out of 7 metrics.

Limitations and Future Work

  • Small Training Set for CNNs: The authors acknowledge that the sample size used for training the CNN was small, and using different tracers and more patients could improve generalization.
  • Generalizability of CNNs: The current CNN might not generalize well to other reconstruction methods or dose levels. Future work should involve more transfer-learning experiments and comparisons with other methods.
  • Clinical Bias: Clinicians may have a bias towards MLEM due to familiarity, potentially influencing subjective metric scores. More exposure to the novel methods might change these perceptions.
  • Future Work: The authors suggest further validation of these methods on different datasets, such as those from epilepsy patients, and exploring other reconstruction methods. They also highlight the potential for economic savings and reduced patient/staff radiation dose with these advanced techniques.

Summary updated

Abstract

Reduced dose Positron Emission Tomography (PET) lowers the radiation dose to patients and reduces costs. Lower count data, however, degrades reconstructed image quality. Advanced reconstruction methods help mitigate image quality losses, but it is important to assess the resulting images from a clinical perspective. Two experienced clinicians assessed four PET reconstruction algorithms for [<sup>18</sup>F]FDG brain data, compared to a clinical standard reference (Maximum-Likelihood Expectation-Maximization (MLEM)), based on seven clinical image quality metrics: global quality rating, pattern recognition, diagnostic confidence (all on a scale of 0-4), sharpness, caudate-putamen separation, noise, and contrast (on a scale between 0-2). The reconstruction methods assessed were a guided and unguided version of self-supervised maximum <i>a posteriori</i> EM (MAPEM) (where the guidance case used the patient's MR image to control the smoothness penalty). For 3 of the 11 patient datasets reconstructed, post-smoothed versions of the MAPEM reconstruction were also considered, where the smoothing was with the point-spread-function used in the resolution modelling. Statistically significant improvements were observed in sharpness, caudate-putamen separation, and contrast for self-supervised MR-guided MAPEM compared to MLEM. For example, MLEM scored between 1-1.1 out of 2 for sharpness, caudate-putamen separation and contrast, whereas self-supervised MR-guided MAPEM scored between 1.5-1.75. In addition to the clinical evaluation, pre-trained Convolutional Neural Networks (CNNs) were used to assess the image quality of a further 62 images. The CNNs demonstrated similar trends to the clinician, showing their potential as automated standalone observers. Both the clinical and CNN assessments suggest when using only 5% of the standard injected dose, self-supervised MR-guided MAPEM reconstruction matches the 100% MLEM case for overall performance. This makes the images far more clinically useful than standard MLEM.
BibTeX
@article{hopson2024clinical,
  author = {Jessica B. Hopson and Sam Ellis and Anthime Flaus and Colm J. McGinnity and Radhouène Neji and Andrew J. Reader and Alexander Hammers},
  title = {{Clinical and Deep-Learned Evaluation of MR-Guided Self-Supervised PET Reconstruction}},
  journal = {IEEE Transactions on Radiation and Plasma Medical Sciences},
  year = {2024},
  doi = {10.1109/trpms.2024.3496779},
}
Figure