Partial Volume Effect

PVE occurs when a voxel straddles the boundary between two or more distinct tissue types. The resulting signal intensity is a weighted average of the constituent tissues' properties, rather than representing any single tissue. This creates a gradual intensity transition at boundaries that should be sharp, effectively blurring the true anatomical edge. The effect is most pronounced in thick-slice acquisitions and at oblique tissue interfaces where the boundary passes diagonally through a voxel.

PVE directly degrades pixel-level classification performance at tissue boundaries. Key consequences include:

• Boundary uncertainty: Classifiers struggle to assign a definitive label to mixed-signal voxels
• Volume underestimation: Small structures may be partially volumed into surrounding tissue, reducing their apparent size
• Dice Score degradation: Boundary errors disproportionately penalize overlap metrics
• False partial volume: Two adjacent structures of different intensities can create an apparent third tissue class at their interface

PVE severity varies by imaging modality and acquisition parameters:

• CT: Predictable linear averaging of Hounsfield Units; a voxel containing 50% soft tissue (40 HU) and 50% bone (400 HU) yields ~220 HU
• MRI: Complex signal averaging due to non-linear tissue properties; partial voluming of gray matter and white matter creates ambiguous voxels at the cortical boundary
• PET/SPECT: PVE causes spill-over of activity from hot regions into adjacent cold regions, leading to quantitative errors in standardized uptake values (SUV)
• Slice thickness is the dominant controllable factor: thinner slices reduce through-plane partial voluming

Several computational approaches address PVE in segmentation workflows:

• Partial volume correction (PVC) algorithms: Methods like the Geometric Transfer Matrix and Müller-Gärtner correction estimate and compensate for tissue mixing
• Sub-voxel segmentation: Techniques that estimate fractional tissue contributions within each voxel rather than assigning a single hard label
• Super-resolution reconstruction: Combining multiple acquisitions to synthesize higher-resolution volumes with reduced PVE
• Loss function design: Soft labeling and boundary-weighted loss functions that account for partial volume uncertainty during training
• Thin-slice acquisition: Reducing slice thickness at scan time remains the most direct physical mitigation

PVE is intimately linked to the imaging system's Point Spread Function (PSF). The PSF describes how a point source is blurred by the imaging system, and PVE is the discrete sampling manifestation of this blur. In MRI, the PSF is determined by k-space sampling and reconstruction filters. In CT, it is influenced by focal spot size, detector element dimensions, and reconstruction kernels. Understanding the PSF allows for deconvolution-based partial volume correction, though this approach is sensitive to noise amplification.

PVE has direct consequences for quantitative imaging biomarkers:

• Tumor volumetry: PVE at tumor margins introduces systematic errors in longitudinal size measurements, potentially affecting RECIST-based response assessment
• Radiomics feature stability: Texture features extracted near boundaries are contaminated by partial volume mixing, reducing their reproducibility
• PET metabolic tumor volume (MTV): Threshold-based segmentation methods are highly sensitive to PVE-induced edge blurring
• Radiotherapy target delineation: PVE contributes to inter-observer variability in Gross Tumor Volume (GTV) contouring, with implications for dose planning margins