Cross-Scanner Harmonization

A foundational intensity normalization technique that transforms the histogram of a source image to match a reference template. How it works: The cumulative distribution function (CDF) of the source is mapped to the CDF of a target template, effectively remapping pixel intensities. Clinical relevance: Reduces gross intensity variations between GE, Siemens, and Philips scanners before feeding images into a downstream segmentation model. Limitations: Purely intensity-based; does not correct structural or textural differences introduced by reconstruction kernels.

A per-image standardization technique that rescales intensities to have zero mean and unit variance. Formula: (x - μ) / σ, where μ and σ are the mean and standard deviation of the image's voxel intensities. Why it matters: Simple, fast, and effective as a first-pass normalization step for most cross-scanner transfer learning pipelines. Caveat: Global statistics can be skewed by large pathologies or background air, so tissue-specific or masked Z-score normalization is often preferred for brain MRI or body CT.

An unpaired image-to-image translation framework that learns to map images from a source scanner domain to a target scanner domain without requiring paired examples. Architecture: Uses a cycle-consistency loss to ensure F(G(source)) ≈ source. Application: Transforms non-contrast CTs from Scanner A to appear as if acquired on Scanner B, harmonizing texture and contrast uptake patterns. Advantage: Operates at the pixel level, directly modifying the image appearance before it enters a diagnostic model.

A feature-level harmonization approach that trains a feature extractor to produce scanner-invariant representations. Mechanism: A gradient reversal layer inserted between the feature extractor and a domain classifier forces the network to learn features that are discriminative for the task but uninformative about the scanner of origin. Benefit: The downstream diagnostic model operates on a harmonized feature space, making it robust to unseen scanner protocols without modifying raw images.

A modality-specific normalization for CT that rescales raw pixel values to standardized Hounsfield Units based on scanner calibration parameters. Key thresholds: Air = -1000 HU, Water = 0 HU, Bone = +400 to +1000 HU. Implementation: Apply a fixed window/level clipping (e.g., [-1000, 400] HU for soft tissue) followed by min-max scaling. Critical step: Ensures that tissue densities are physically meaningful and consistent across vendors before any deep learning pipeline.

A test-time adaptation technique that updates the running mean and variance statistics of a pre-trained model's BatchNorm layers using target domain data. Process: Pass unlabeled target images through the model in inference mode to recompute channel-wise statistics, adapting to the new scanner's intensity distribution. Advantage: Requires no backpropagation or label access; a lightweight, inference-only harmonization step. Use case: Deploying a model trained on Siemens data directly onto GE scanner outputs in a clinical setting.

A lightweight test-time adaptation technique that updates the running mean and variance statistics of a model's Batch Normalization layers using target domain data. This approach combats covariate shift without requiring any labeled examples or gradient-based fine-tuning.

• Computationally negligible overhead at inference
• Adapts to each new scanner on-the-fly
• Often used as a baseline for more complex harmonization methods