-
Notifications
You must be signed in to change notification settings - Fork 1
Standardized features computation
All standardized features are computed in MNI space after affine and deformable registration (SyNQuick from ANTs), and can be generated for any segmented structure (i.e., tumor core, contrast-enhancing rest tumor, FLAIR hyperintensity).
-
The volume of the structure, expressed in ml, is computed as:
$\frac{\sum_{x=1}^n v_x = s}{\sum_{x=1}^n v_x = b} * (s_x * s_y * s_z) * 1e^{-3}$ , where$v_x = s$ represents voxels belonging to the structure and$v_x = b$ voxels belonging to the brain, and$s_x * s_y * s_z$ are the spacing values along each axis representing the physical size of a voxel. -
The structure to brain ratio, following the previous equation and expressed in %, is computed as:
$\frac{volume_{structure}}{volume_{brain}} * 100$
- The long-axis diameter, expressed in mm, is computed as the length of the major axis of the ellipse that has the same normalized second central moments as the region.
- The short-axis diameter, expressed in mm, is computed as the length of the minor axis of the ellipse that has the same normalized second central moments as the region.
- The Feret diameter, expressed in mm, is computed as the longest distance between points around a region's convex hull contour.
- The equivalent area, expressed in mm, is computed as the diameter of the circle with the same area as the region.
As preprocessing, a cleaning step is performed by running a connected component analysis and removing structure components that are smaller than 0.1 ml, to avoid potential segmentation noise interfering with the computation.
- Status, expressed as boolean, is set to true if at least two structure components exist with a minimum distance between them of 5 mm.
- Elements, expressed as number, is the total amount of identified structure components.
- Distance, expressed in mm, is computed as the maximum distance between the largest component and any other component, using the Hausdorff95 computation.
A brain lateralization mask, based on the MNI atlas is used to identify left and right brain hemispheres.
- Status, expressed as boolean, is set to true if both hemispheres are overlapped by the structure.
- Right hemisphere percentage, expressed in %, is computed as
$\frac{\sum_{x=1}^n (v_x = s) \cup (v_x = b_r)}{\sum_{x=1}^n v_x = s} * 100$ - Left hemisphere percentage, expressed in %, is computed as
$\frac{\sum_{x=1}^n (v_x = s) \cup (v_x = b_l)}{\sum_{x=1}^n v_x = s} * 100$ - Minimum, maximum, and mean depths, expressed in mm, are measured by computing the exact Euclidean distance between the structure and the boundaries of the brain.
The resectability maps, as presented in this article, capture the surgical treatment decision to stop glioma removal and allow quantification of the extent of resection per voxel. As a map has been computed for each brain hemisphere independently, the map selected for computing the resectability features (RMAP) is the map of the hemisphere most overlapped by the structure (i.e., > 50%). OBS: Those features only make sense when the structure is a tumor core segmentation.
- Resectable volume, expressed in ml,
$\sum_{x=1}^n RMAP(v_x = s) * 1e^{-3}$ - Residual volume, expressed in ml,
$(\sum_{x=1}^n (v_x = s) - \sum_{x=1}^n RMAP(v_x = s)) * 1e^{-3}$ - Average resectability, expressed in ,
$\frac{\sum_{x=1}^n RMAP(v_x = s)}{\sum_{x=1}^n v_x = s}$
Multiple cortical structures profiles, all expressed in MNI space, are available for computing the overlap and distance between each structure of the atlas and the segmented structure of interest.
- MNI: frontal lob (left/right, wm/gm), temporal lob (left/right, wm/gm), parietal lob (left/right, wm/gm), occipital lob (left/right, wm/gm), cerebellum (left/right), brain stem, lateral ventricle (left/right), third ventricle, fourth ventricle, extracerebral CSF, caudate (left/right), putamen (left/right), thalamus (left/right), subthalamic nucleus (left/right), globus pallidus (left/right), fornix (left/right).
- Schaefer7: visual, somatomotor, dorsal attention, salience ventral attention, limbic, frontoparietal control, default.
- Schaefer17: visual central, visual peripheral, somatomotor (A/B auditory), dorsal attention (A/B), salience ventral attention (A/B), limbic (orbitofrontal/temporopolar), frontoparietal control (A/B/C), default (A/B/C), temporoparietal.
- Hardvard-Oxford: frontal pole, insular cortex, superior frontal gyrus, middle frontal gyrus, inferior frontal gyrus (triangularis/opercularis), precentral gyrus, temporal pole, superior temporal gyrus (anterior/posterior), middle temporal gyrus (anterior/posterior/temporooccipital), inferior temporal gyrus (anterior/posterior/temporooccipital), postcentral gyrus, superior parietal lobule, supramarginal gyrus (anterior/posterior), angular gyrus, lateral occipital cortex (inferior/superior), intracalcarine cortex, frontal medial cortex, juxtapositional lobule cortex, subcallosal cortex, paracingulate cortex, cingulate gyrus (anterior/posterior), precuneous cortex, cuneal cortex, frontal orbital cortex, parahippocampal gyrus (anterior/posterior), lingual gyrus, temporal fusiform cortex (anterior/posterior), temporal occipital fusiform cortex, occipital fusiform gyrus, frontal operculum cortex, central opercular cortex, parietal operculum cortex, planum polare, Heschls gyrus, planum temporale, supracalcarine cortex, occipital pole.
For each atlas cortical structure, distance and overlap with the segmented structure mask are computed.
- Overlap, expressed in % of the segmented structure volume, is computed as
$\frac{\sum_{x=1}^n (v_x = s) \cup (v_x = a_s)}{\sum_{x=1}^n v_x = s} * 100$ , whereby$v_x = s$ are segmented structure voxels and$v_x = a_s$ are atlas structure voxels. - Distance, expressed in mm, is measured by computing the exact Euclidean distance between the segmented structure and the boundaries of the atlas structure.
Multiple subcortical structures profiles, all expressed in MNI space, are available for computing the overlap and distance between each structure of the atlas and the segmented structure of interest.
- Brain Connectivity and Behaviour (BCB):
- Brain-Grid:
For each atlas subcortical structure, distance and overlap with the segmented structure mask are computed in the same fashion as previously described.
A novel radiological classification system for cerebral gliomas, as introduced in this article. A total of 48 areas are defined (named Brain-Grid Voxels) according to three coordinates (Axial, Coronal, and Sagittal) and four numbers (1-4). OBS: Those features only make sense when the structure is a tumor core segmentation.
- Infiltration, expressed as a number, indicates the amount of Brain-Grid Voxels infiltrated by the structure.
- Overlap, expressed as a % of the segmented structure volume, is computed for each Brain-Grid Voxel using the same formula as described previously.