Hounsfield Unit (HU) Normalization

Normalization maps raw pixel values to standard Hounsfield Units (HU), where water is 0 HU and air is -1000 HU. This allows models to focus on specific tissue density ranges:

• Soft tissue window: -125 to +225 HU for brain, liver, and organs
• Lung window: -1000 to -200 HU for pulmonary parenchyma
• Bone window: +300 to +1500 HU for osseous structures

By clipping intensities to a tissue-relevant range, the model ignores irrelevant density information and reduces the feature space dimensionality.

Raw CT intensities are vendor-specific and depend on acquisition parameters like tube current and voltage. A given tissue can appear at different raw values across GE, Siemens, or Philips scanners.

HU normalization converts these arbitrary values to a standardized physical scale based on linear attenuation coefficients. This ensures a model trained on data from one hospital generalizes to scans from another without catastrophic performance degradation.

After HU rescaling, a common secondary step is z-score normalization:

• Subtract the mean intensity of the volume or region of interest
• Divide by the standard deviation

This produces a distribution with zero mean and unit variance, which stabilizes gradient flow during training and accelerates convergence. The formula is: (x - μ) / σ where μ and σ are computed per-volume or per-dataset.

An alternative to z-score normalization is min-max scaling, which linearly transforms HU values to a fixed range:

• Formula: (x - min) / (max - min) for [0,1] range
• Clipping first: Apply a tissue window (e.g., -1000 to +400 HU) before scaling to avoid outlier pixels dominating the range

This is particularly useful when the model expects input tensors in a specific numerical range, such as sigmoid-activated output layers or pre-trained ImageNet weights adapted for medical imaging.

CT scans can be non-contrast, arterial phase, venous phase, or delayed phase, each with different HU distributions for the same anatomy. A robust normalization pipeline must account for this:

• Metadata-driven normalization: Use DICOM header tags to identify contrast phase and apply phase-specific windows
• Adaptive normalization: Compute statistics per-phase during training to learn invariant representations

Ignoring contrast phase leads to distribution shift and unreliable model predictions in clinical deployment.

Raw CT volumes often contain extreme outliers from metal implants, motion artifacts, or beam hardening. These can distort normalization statistics.

Robust normalization techniques include:

• Percentile clipping: Clip intensities at the 1st and 99th percentiles before scaling
• Foreground masking: Compute statistics only on non-air voxels (HU > -800)
• Median-based scaling: Use median and interquartile range instead of mean and standard deviation

These methods prevent a single metallic hip implant from corrupting the entire volume's normalization.