Head preprocessing¶

PRESTUS preprocesses layered segmentations to set up simulation grids.

preprocessing

Orientation¶

PRESTUS initially creates a grid by orienting the segmentation to the focal plane of the first transducer. In multi-transducer setups, alignment is done to the first specified transducer. Specifically, preproc_align_to_focal_axis performs a two-step rotation to align an arbitrary focal axis (defined by transducer and focus positions) with the z-axis of the coordinate system. This prepares subject-specific head models for ultrasound neuromodulation simulations by standardizing the acoustic propagation direction along z.

Scaling¶

The 3D planning image is then resliced to the requested grid resolution by isotropic scaling and centered regridding. These steps are applied to the planning image, the segmentation, and the bone mask (or (p)CT, if specified). For masks, nearest neighbor interpolation is used, for continuous values (e.g., from a (pseudo-)CT, linear interpolation is applied).

Medium Mapping (incl. segmentation smoothing)¶

Segmentations are mapped onto the requested layer medium masks (to later assign acoustic properties). The standard (10-tissue) SimNIBS 4 segmentation will be converted into the (by default 5) specified layers (controlled by parameters.layers). If a segmentation for the requested layer is not available, requested layers will be removed from the specification. All segmentation IDs that are not explicitly assigned in dedicated layers are automatically included in a baseline water layer. Medium maps contain layer indices corresponding to the respective layer label position in the medium properties (parameters.medium). This allows for flexibility in requested layers with stable medium ids (unless parameters.medium is edited).

Each layer is smoothed using one of the algorithms below. The size of smoothing kernels refer to the grid size of the medium mask, NOT the original dimensions of the planning image. If a multi-layer skull model is requested, both cortical and trabecular layers are initially combined into a skull segment and smoothed. Subsequently, the trabecular bone layer is smoothed and added.

Multiple smoothing algorithms are available:
- gaussian (default) uses imgaussfilt / imgaussfilt3 for isotropic blurring, yielding smooth gradients and moderate edge erosion (default: 1 mm FWHM kernel, converted to voxels at runtime) - box applies a uniform convolution, which is fastest, but also blocky and prone to skull holes at low binarization thresholds (default: 1 mm kernel) - off disables smoothing entirely, passing the raw segmentation mask through without any filtering or binarization

Different binarization thresholds can be defined for skull (default: 50%) and other tissues (default: 50%).

PRESTUS attempts to enforce skull continuity (i.e., absence of holes) either by applying a 'rubberwrap' skull inflation algorithm (default) or 'imclose' to the skull layer. This is controlled by headmodel.skull_fill_method. Potential holes are converted to (cortical) skull.

Cropping¶

To increase computational efficiency, the grid is tightly cropped around the head. The cropping factor is determined by the edges of the layered medium and the location of the transducer. [Prior to adding the transducer, the grid can optionally be expanded via parameters.headmodel.head_pad_mm, which applies symmetric padding and can avoid that the transducer ends out of grid bounds. This should not impact later computational efficiency of the simulation, as regions beyond the transducer + PML will afterwards be cropped]. To guide the bounds of the head area of interest in the simulation, the CSF segmentation (if available) will be expanded by parameters.headmodel.csf_expansion [in mm, converted to voxels at the current grid resolution] to guide bounds of the layered medium. Areas outside of the crop mask will be assigned to the water layer (and are eligible for cropping). Tight crop dimensions are then defined by enclosing the layered medium mask, the transducer geometry, plus a PML buffer. This is followed by FFT-optimized resizing via find_min_factor, which expands dimensions up to parameters.grid.max_expand voxels for faster FFT convolution. The medium is anchored at the original origin (transducer/target unchanged relative to head), with symmetric expansion/padding around the content to fit FFT requirements while maintaining anatomical alignment for k-Wave propagation.

All image transformations are concatenated to update transducer and target positions.

Acoustic property mapping¶

See acoustic property mapping.

Smoothing of acoustic property maps¶

Sharp transitions in the acoustic properties of adjacent media can result in substantial wave reflections. To smoothen tissue transitions, acoustic property maps can be smoothed prior to the simulation. This is governed by parameters.headmodel.smooth_properties. Identical smoothing will be applied to all acoustic property maps. This smoothing is deactivated by default, in part to allow users to set the desired smoothing kernel at the intended grid size. If set to true, this will apply the same smoothing parameters as used during the creation of the medium maps (headmodel.smooth_method and headmodel.smooth_fwhm_mm, default: 1 mm FWHM gaussian).

pCT. If pCTs are generated using PRESTUS, they automatically have smoothing applied at the skull edges (see the pCT documentation). If acoustic properties are mapped from pCTs, smoothing of the resulting acoustic property maps may not be intended. Recommended settings are an active area of investigation.