Voxel

A voxel (volume element) is a value on a regular grid in 3D space, analogous to a pixel in 2D. It represents a discrete sample of a continuous volume. In CT imaging, the voxel value is a Hounsfield Unit (HU) quantifying radiodensity. In MRI, it represents signal intensity. The voxel is the fundamental data structure for all volumetric operations, from Filtered Back Projection to Deep Learning Reconstruction.

A voxel's dimensions are defined by its in-plane pixel spacing and slice thickness. An isotropic voxel has equal dimensions on all axes (e.g., 1mm x 1mm x 1mm), forming a perfect cube. Anisotropic voxels are non-cubic, typically with a larger slice thickness than pixel spacing. Isotropic volumes are preferred for Multi-Planar Reconstruction (MPR) as they prevent distortion when reslicing the dataset along arbitrary planes.

The Partial Volume Effect is a critical artifact where a single voxel contains a mixture of multiple tissue types. The resulting signal is a weighted average of the constituent materials, causing blurred boundaries. This is especially problematic for small structures like thin bone or small vessels. Mitigation strategies include acquiring thinner slices to reduce voxel size or using advanced segmentation models like nnU-Net that are trained to handle boundary uncertainty.

Raw voxel values, such as Hounsfield Units, often span a wide range (e.g., -1000 for air to +3000 for dense bone). Windowing is the non-linear mapping of this range to display grayscale values. A window width defines the range of values displayed, and a window level sets the center. This operation does not alter the underlying voxel data but is essential for visualizing specific tissues, like a narrow lung window or a wide bone window.

In deep learning, a 3D medical scan is loaded as a multi-dimensional array or tensor. A typical volume has the shape (Depth x Height x Width) or (D x H x W). For a DICOM Series, the depth corresponds to the number of slices. Architectures like the 3D U-Net process these volumetric tensors directly, using 3D convolutions to learn spatial features across all three axes, preserving inter-slice context that 2D models miss.

An isosurface is a 3D surface representing points of a constant value within a volume of voxels. The Marching Cubes algorithm extracts a polygonal mesh by iterating through the voxel grid, determining how an isosurface intersects each voxel. This converts a voxel-based segmentation mask into a geometric model for Surface Rendering (SR) or 3D printing, enabling visualization of a specific anatomical structure's boundary.