Maximum Probability Atlas Implementation in PNEURO

<< Click to Display Table of Contents >>

Navigation:  Brain VOIs Based on Maximum Probability Atlas >

Maximum Probability Atlas Implementation in PNEURO

In PNEURO the use of the N30R83 atlas (and the other atlases with PET and MR normalization templates) is supported in four situations, namely studies with PET and T1-MR, studies with a functional and an "anatomical" PET, PET-only studies, and MR-only studies. The corresponding workflows are outlined in the following.

Analysis of Study with PET and T1-weighted MRI

This is the most accurate workflow. The other workflows are essentially adapted subsets.

1.Loading of the PET image series which may be static or dynamic.

2.Dynamic PET case: Averaging of the PET series in a specified acquisition range. The averaged PET image is used in the following for all steps except for the statistics calculation.

3.Loading of the T1-weighted MR image series.

4.Calculation of the individual gray matter probability map by segmentation of the MR image. This only possible if the atlas definition includes tissue probability maps.

5.Splitting of the brain into the left and right hemispheres as well as cerebellum. This step is relevant for separating the white matter parts correspondingly.

6.Rigid matching of the PET image to the MR image and interactive visual assessment of the alignment by the user.

7.Spatial normalization of the MR image to the MNI T1 template and interactive visual assessment of the alignment by the user. The normalization according to tissue probability maps and subsequent masking of atlas segments according to gray matter probability is widely successful in cortical regions, but can be insufficient for deep nuclei around enlarged ventricular spaces. PMOD introduces a hybrid-AI solution: it utilizes the trusted methods for cortical gray matter segmentation, but allows replacing the deep nuclei segments with the result of prediction with a trained neural network.

8.Transformation of the label atlas to the MR space and display of the result as an overlay on the MR image.

9.Intersection of the cortical structures with the gray matter probability map at a user-defined probability level.

10.Calculation of the outline contours of the masked structures. They are presented in a VOI editor together with the MR images, so that the user can adjust them interactively and save the final VOI set.

11.Application of the VOIs to the matched PET series for calculating statistics. This results in TACs in the case of a dynamic PET series, and simple statistics otherwise. Optionally, a partial-volume correction can be applied during the statistics calculation.

12.Dynamic PET case: The resulting TACs can directly be transferred to the kinetic modeling tool. Alternatively, pixel-wise models can be applied for parametric mapping.

Note that instead of defining the VOIs in the MR space as described above, they may be defined in the atlas or the PET space, and the statistics calculated after transforming the PET series into the selected space.

Analysis of Study with Functional and Anatomical PET

If no T1-weighted MRI is available, but an additional PET with more anatomical information (e.g. FDG), the role of the MRI for the normalization can be taken over by the "anatomical" PET (called "FDG PET" in the following).

1.Loading of the PET image series which may be static or dynamic.

2.Dynamic PET case: Averaging of the PET series in a specified acquisition range. The averaged PET image is used in the following for all steps except for the statistics calculation.

3.Loading of the FDG PET image series.

4.Rigid matching of the two PET images, and interactive visual assessment of the alignment by the user.

5.Spatial normalization of the FDG PET image to the MNI PET template and interactive visual assessment of the alignment by the user. Optionally, a user-defined normalization template may be used instead of the standard MNI PET template.

6.The label atlas is transformed to the subject space and shown as an overlay on the PET image.

7.The cortical structures in the transformed label atlas can be intersected with a standard gray matter probability map (if available for the atlas) which has been transformed to the subject space. The user can define the probability level used for masking, and whether the non-cortical structures are masked or not.

8.The structures resulting from transforming and masking are outlined and shown in a VOI editor, so that the user can adjust them interactively and save the final VOI set.

9.The VOIs are applied to the PET series for calculating statistics. This results in TACs for a dynamic PET series, and simple statistics otherwise. Optionally, a partial-volume correction can be applied during the statistics calculation.

10.Dynamic PET case: The resulting TACs can directly be transferred to the kinetic modeling tool, or alternatively pixel-wise models applied for parametric mapping.

As an alternative to the workflow described above which performs the calculations in the individual's anatomy on the original PET images, the PET images can be transformed to the atlas space and all calculations performed in analogy.

Analysis of PET-only Study

If an anatomical image series is lacking the processing sequence reduces to the following steps:

1.Loading of the PET image series which may be static or dynamic.

2.Dynamic PET case: Averaging of the PET series in a specified acquisition range. The averaged PET image is used in the following for all steps except for the final statistics calculation.

3.Spatial normalization of the PET image to the atlas PET template. The user has to visually check that the transformed PET image is in reasonable spatial alignment with the template. If this is the case, the normalization transform establishes a bidirectional mapping between the space of the subject and the template. Optionally, a user-defined normalization template may be used instead of the standard atlas PET template.

4.The label atlas is transformed to the subject space and shown as an overlay on the PET image.

5.The cortical structures in the transformed label atlas can be intersected with a standard gray matter probability map (if available for the atlas) which has been transformed to the subject space. The user can define the probability level used for masking, and whether the central structures are masked or not.

6.The structures resulting from transforming and masking are outlined and shown in a VOI editor, so that the user can adjust them interactively and save the final VOI set.

7.The VOIs are applied to the PET series for calculating statistics. This results in TACs for a dynamic PET series, and simple statistics otherwise. Optionally, a partial-volume correction can be applied during the statistics calculation.

8.Dynamic PET case: The resulting TACs can directly be transferred to the kinetic modeling tool, or alternatively pixel-wise models applied for parametric mapping.

As an alternative to the workflow described above which performs the calculations in the individual's anatomy on the original PET images, the PET images can be transformed to the atlas space and all calculations performed in analogy.

Analysis of MR-only Study

In this case the workflow reduces to the following steps:

1.Loading of the T1-weighted MR image series.

2.Calculation of gray matter probability maps by segmentation of the MR image.

3.Spatial normalization of the MR image to the atlas T1 template  and interactive visual assessment of the alignment by the user. The normalization according to tissue probability maps and subsequent masking of atlas segments according to gray matter probability is widely successful in cortical regions, but can be insufficient for deep nuclei around enlarged ventricular spaces. PMOD introduces a hybrid-AI solution: it utilizes the trusted methods for cortical gray matter segmentation, but allows replacing the deep nuclei segments with the result of prediction with a trained neural network.

4.Transformation of the label atlas to the MR space and display of the result as an overlay on the MR image.

5.Intersection of the cortical structures with the gray matter probability map at a user-defined probability level.

6.Calculation of the outline contours of the masked structures. They are presented in a VOI editor together with the MR images, so that the user can adjust them interactively and save the final VOI set.

7.Application of the VOIs for calculating statistics. In the absence of PET the statistics are reduced to the calculation of the VOI volume.

Documentation of the Workflows

The implementation of the workflows described above in PNEURO is very similar. Therefore, only the first workflow with PET and a T1-weighted MRI will be described in full detail, while the others are restricted to the essential parts. Please refer to the first workflow description if any questions arise.