Work Package 4

MR assisted IDIF on MR/BrainPET

With the advances in PET and PET/MR technology allowing higher quantification accuracy, dynamic PET studies using modeling for full quantification are becoming more prevalent. The gold standard requires arterial blood sampling to estimate the arterial input function (AIF). However, this method can be complex and is invasive due to arterial cannulation. Conversely, the image-derived input function (IDIF) method is a non-invasive alternative enabling estimation with no or minimal blood sampling. This work aimed to develop a fully automated method for IDIF estimation in PET/MR and dedicated PET studies, especially where no anatomical information is available.

Our method was implemented with Python, comprising the following steps:

  1. Open MR and PET images in Nifti format
  2. Creates an average image of the dynamic PET
  3. Perform MR bias correction in T1 MPRAGE
  4. Rigid co-registration of PET average image to T1 MPRAGE bias corrected image
  5. Save transformation matrices for further needs
  6. Remove skull from T1 MPRAGE MR bias corrected image
  7. Apply median filter to T1 MPRAGE MR corrected image without skull
  8. Select internal carotids (ICA) in MR image using probability map created from atlas
  9. Binary mask created from segmented ICA
  10. Dynamic (for each frame) binary ICA mask created
  11. Partial volume correction (PVC) was applied to PET dynamic image
  12. Dynamic ICA mask applied to PET dynamic image
  13. Select and average the 4 hottest5 pixels values per plane
  14. Get time from PET framing scheme
  15. Get arterial blood curve information (gold standard) from the study for comparison/validation
  16. Plot image-derived and arterial blood curves
  17. Fit blood curves with Gaussian
  18. Calibrate IDIF curve based on AIF curve
  19. Calculate the area under the curve (AUC) for the peak (P) and tail (T) for validation – ratio between IDIF and AIF AUCs

The developed automated tool appears to be feasible for PET/MR and dedicated PET applications and the need for interaction/manual operations from the user is minimal – only the input image path and blood data file need to be changed.

The segmentation using the brain vessel atlas had yet some limitations. Notably, the vessel atlas used for creating the ICA probability maps only contained the upper part of the ICA, resulting in an incomplete representation that disregarded individual anatomical variations. Ongoing assessment aims to determine the influence of different segmentation methods, partial volume correction methods, and fitting functions. Further evaluation of the general robustness of the method and its performance with other radiotracers and in the presence of volunteer motion will be made, including the test of other PVC methods available for PET image correction and the evaluation of different fit functions.

In any case and in summary, the project has succeeded in optimizing the determination of the IDIF, ensuring that it provides a good approximation. This is a significant accomplishment, especially in applications where precision is essential. A reliable IDIF contributes to the overall robustness and reliability of the project and its dual focus on user-friendliness and IDIF accuracy demonstrates a well-rounded approach, addressing both usability and functionality.

Work Package 1

UMC-PET: a fast and flexible Monte Carlo PET simulator

The GPU-based Ultra-fast Monte Carlo positron emission tomography simulator (UMC-PET) incorporates the physics of the emission, transport and detection of radiation in PET scanners. It includes positron range, non-colinearity, scatter and attenuation, as well as detector response. The objective of this work within Work Package 1 is to present and validate UMC-PET as a a multi-purpose, accurate, fast and flexible PET simulator. We compared UMC-PET against PeneloPET, a well-validated MC PET simulator, both in preclinical and clinical scenarios. Different phantoms for scatter fraction (SF) assessment following NEMA protocols were simulated in a 6R-SuperArgus and a Biographm MR scanner, comparing energy histograms, NEMA SF, and sensitivity for different energy windows. UMC-PET employs a voxelized scheme for the scanner, patient adjacent objects (such as shieldings or the patient bed), and the activity distribution. This makes UMC-PET extremely flexible. Its high simulation speed allows applications such asMCscatter correction, faster SRM estimation for complex scanners, or even MC iterative image reconstruction.

In our workflow above, all the inputs regarding computational performance and physical simulation are read. The Positron Range kernel is executed prior to the particle simulation. The photon simulation consists of four routines: photon generation, particle tracking, singles processing, and coincidence sorting. List files and cumulative histograms are generated during the simulation procedure. In the image below, a Voxelized image of the mMR scanner with the inner coil inside the PET bore and a pelvis CT segmented in adipose tissue, bone and water is shown.

As a conclusion, our UMC-PET simulatorhas proven to be a fast, versatile and accurate Monte Carlo code for PET simulation with GPU. UMC-PET has been developed with a primary focus on enhancing image reconstruction, including scatter correction and SRM estimation, as well as supporting the design of PET scanners. Its approach to define scanners and detectors in a voxelized manner simplifies its application and makes it possible to consider various geometries and the majority of detector configurations currently in use or planned for. The incorporation of precomputed tables for scattered photons and attenuation coefficients, along with a factorized scheme for the main physic principles underlying the technique, has simplified the code without adversely affecting simulation accuracy when compared to other approaches.