Isotropic Resampling

Transforms voxels into perfect cubes where spacing is identical in all three axes (e.g., 1.0mm × 1.0mm × 1.0mm). This eliminates directional bias in convolutional kernels, allowing 3D filters to learn rotationally invariant features regardless of the original acquisition plane. Without this step, a model would treat the superior-inferior axis differently than the axial plane, degrading segmentation accuracy for structures like the aorta or spine.

The resampling process requires selecting an interpolation algorithm that balances accuracy and computational cost:

• Nearest Neighbor: Preserves discrete label values; mandatory for resampling segmentation masks to avoid introducing artificial classes.
• B-Spline (Cubic): Preferred for intensity images (CT/MRI) as it produces smooth, continuous values with minimal aliasing artifacts.
• Linear: A faster alternative to B-Spline, offering acceptable quality for many applications.
• Windowed Sinc: The theoretical gold standard using a Lanczos kernel, but computationally expensive for large 3D volumes.

Clinical CT and MRI scans are typically acquired with high in-plane resolution (e.g., 0.5mm × 0.5mm) but thick slice spacing (e.g., 3.0–5.0mm). This creates elongated voxels that distort anatomical geometry. Isotropic resampling corrects this by interpolating new slices between acquired ones, effectively synthesizing a continuous volume. The degree of correction is defined by the resampling factor: the ratio of original slice thickness to target isotropic spacing.

Deep learning architectures like 3D U-Net and nnU-Net extract cubic patches from volumes during training. If the underlying voxels are anisotropic, a 'cubic' patch represents a rectangular physical region, causing the network to learn distorted spatial relationships. Isotropic resampling ensures that each patch covers a consistent physical volume, enabling the model to learn true 3D anatomical context—critical for segmenting small, tortuous structures like coronary arteries or cortical sulci.

The nnU-Net framework automates architecture selection based on the resampled median voxel spacing of a dataset. This 'spacing-aware' design adjusts kernel sizes, pooling operations, and patch dimensions to match the physical scale of anatomical structures. For isotropic data with fine spacing (e.g., 0.5mm), nnU-Net uses smaller kernels and larger patch sizes to capture detailed texture. For coarser spacing (e.g., 2.0mm), it prioritizes larger receptive fields to capture global context.

Resampling to a fine isotropic grid (e.g., 0.5mm) dramatically increases the total voxel count—a 512×512×300 volume at 1.0mm spacing becomes a 1024×1024×600 volume at 0.5mm, an 8× increase in memory. This directly impacts GPU VRAM requirements and training time. Practical implementations often choose a target spacing that balances anatomical fidelity with hardware constraints, typically matching the median in-plane resolution of the original acquisition.

A voxel is a volumetric pixel representing a value on a 3D grid. Unlike a 2D pixel, a voxel has explicit depth, height, and width dimensions. In anisotropic medical scans (e.g., CT with 0.5mm x 0.5mm in-plane but 3.0mm slice thickness), voxels are elongated rectangular prisms. Isotropic resampling transforms these into perfect cubes, ensuring a 1mm feature in the X-axis has the same spatial meaning as a 1mm feature in the Z-axis. This uniformity is critical for 3D convolutional kernels that assume equal spatial significance in all directions.

Resampling requires estimating intensity values at new grid positions. Common techniques include:

• Nearest Neighbor: Assigns the value of the closest original voxel. Preserves label boundaries; essential for resampling segmentation masks.
• Trilinear Interpolation: Computes a weighted average of the 8 nearest voxels. Smooth and fast; standard for CT/MRI intensity images.
• B-Spline Interpolation: Uses higher-order basis functions for smoother results. Preferred for high-resolution isotropic reconstruction but computationally more expensive.
• Lanczos Interpolation: A windowed sinc filter offering excellent frequency preservation, often used in research-grade preprocessing.

Anisotropy in medical volumes stems from acquisition physics and clinical pragmatism. CT scanners generate high in-plane resolution (0.3-0.7mm) but thicker slices (1-5mm) to reduce radiation dose and scan time. MRI sequences often have large inter-slice gaps to manage acquisition duration. 3D CNNs like the U-Net architecture expect isotropic input; feeding anisotropic data causes the network to learn distorted spatial features, degrading segmentation accuracy for small structures like vessels or lesions that span only a few slices.

The partial volume effect occurs when a single voxel contains multiple tissue types, producing a blended intensity value. In thick-slice anisotropic data, this effect is severe at tissue boundaries. Isotropic resampling cannot recover lost information, but it standardizes the representation so the segmentation network learns consistent boundary profiles. Advanced resampling pipelines may incorporate super-resolution techniques to synthetically increase through-plane resolution, but standard interpolation only redistributes existing information—it does not create new detail.