Temporal Comparison

A non-rigid spatial alignment technique that warps a prior mammogram to precisely match the current exam's geometry. Unlike rigid registration, which only applies translation and rotation, deformable registration uses B-spline or diffeomorphic transformations to account for tissue compression differences, patient positioning variations, and natural anatomical changes between exams. The algorithm optimizes a similarity metric—typically mutual information or normalized cross-correlation—to generate a dense deformation field mapping every pixel from the prior to the current view.

A derived image computed by subtracting the registered prior mammogram from the current exam on a pixel-by-pixel basis. In the resulting difference image, stable anatomy cancels out to a neutral gray background, while new or growing tissue appears as bright or dark regions. This technique makes interval change visually conspicuous, revealing:

• New microcalcification clusters that were absent previously
• Developing asymmetries representing tissue proliferation
• Enlarging mass margins indicating potential tumor growth

A coarse alignment step that identifies stable anatomical reference points—such as the nipple profile, pectoral muscle boundary, and skin line contour—to establish an initial transform before fine deformable registration. This hierarchical approach prevents the optimization algorithm from converging on local minima caused by large initial displacements. Key landmarks include:

• Nipple location: A consistent anterior reference point
• Chest wall interface: The posterior boundary in MLO views
• Vascular patterns: Unique branching structures that serve as internal fiducials

A preprocessing step that standardizes the grayscale distribution between current and prior mammograms to compensate for acquisition parameter drift over time. Without normalization, differences in kVp settings, compression force, or detector calibration between exams can produce spurious signals in the subtraction image. Techniques include histogram matching to a reference distribution and z-score normalization of pixel intensities within the breast region, ensuring that only true biological change survives the subtraction process.

A color-coded overlay that maps the statistical significance of temporal change onto the current mammogram for rapid radiologist interpretation. Rather than displaying raw subtraction values, the system computes a localized change score using:

• Z-score maps: Pixel-wise deviation from expected noise distribution
• Region growing: Clustering contiguous pixels exceeding a significance threshold
• Temporal context: Weighting change magnitude against known benign fluctuation patterns Hot colors indicate high-confidence interval change requiring focused review.

A verification step that cross-references detected temporal changes against the contralateral breast to suppress benign symmetric developments. Physiological changes such as hormonal density fluctuations or bilateral stromal involution often manifest symmetrically across both breasts. By comparing the left-right asymmetry of any detected interval change, the system can downgrade findings that appear in both breasts simultaneously—a pattern strongly associated with benign processes rather than unilateral malignancy.

The foundational preprocessing step that spatially aligns a current mammogram with a historical one using rigid or deformable transformations. Without accurate registration, temporal subtraction produces meaningless artifacts. Techniques range from intensity-based alignment using mutual information metrics to landmark-based registration that anchors on anatomical fiducials like the nipple, pectoral muscle boundary, and vascular patterns. Deformable registration further compensates for differential tissue compression between exams, ensuring that corresponding fibroglandular structures are mapped to the same coordinate space before subtraction.

A computational technique that pixel-wise subtracts a registered prior mammogram from a current one, generating a difference image that highlights interval change. Stationary normal anatomy cancels out, while new or evolving lesions appear as residual signal. This method is particularly effective for detecting architectural distortions and asymmetric densities that develop slowly over screening intervals. Advanced implementations use non-linear warping to correct for local tissue deformation before subtraction, minimizing background clutter and improving the conspicuity of subtle malignancies.

The algorithmic identification of statistically significant differences between temporally separated mammograms. Modern deep learning approaches use Siamese neural networks that process both exams through shared-weight encoders, producing feature embeddings that are compared via distance metrics or attention mechanisms. Key targets include:

• New calcification clusters appearing in previously clear tissue
• Enlarging mass margins indicating active growth
• Developing asymmetry where fibroglandular tissue density increases unilaterally
• Architectural distortion progression with more pronounced radiating lines

A specialized neural network design where two identical subnetworks process the current and prior mammograms in parallel, sharing weights to ensure consistent feature extraction. The twin outputs are fused through a distance layer—commonly using contrastive loss or triplet loss—that learns to minimize the embedding distance for unchanged tissue while maximizing it for true pathological change. This architecture is inherently robust to variations in positioning and compression because it learns invariant representations of anatomical structures rather than pixel-level correspondence.

Non-pathological signals in subtraction images that can mimic or obscure true interval change. Common sources include:

• Misregistration errors from imperfect spatial alignment, creating bright rims at tissue boundaries
• Compression variability causing differential tissue spreading between exams
• Skin fold changes that appear as new linear densities
• Positioning differences in the pectoral muscle or inframammary fold Robust temporal comparison systems incorporate artifact classification modules trained to distinguish these benign subtraction residuals from genuine suspicious findings, directly reducing false positive recall rates.

An extension of temporal comparison that analyzes multi-timepoint trajectories rather than just two exams. By tracking the evolution of tissue patterns across a patient's entire screening history, these models learn personalized baseline representations of normal parenchymal change. A deviation from this individualized trajectory—such as an accelerating rate of density increase or a calcification cluster that grows faster than the patient's historical norm—triggers a high-suspicion flag. This approach reduces the confounding effect of naturally dense or changing breast tissue.