Tracking anatomical changes over time—such as the progression of brain atrophy in dementia or the response of a tumor to treatment—is a cornerstone of longitudinal medical imaging. However, these changes are complex, occurring across multiple spatial scales simultaneously. This paper, presented at MICCAI 2013, introduces a robust probabilistic framework to automatically decompose deformation fields into a "sparse scale-space," allowing for the precise localization of significant volume changes.
The Challenge of Multi-Scale Anatomical Change
Anatomical deformations are rarely uniform. In Alzheimer’s disease, for instance, clinicians observe both localized thinning of the gray matter (measured in millimeters) and broader regional loss in structures like the hippocampus and ventricles. Existing analytical methods often struggle with this variance. Local measures, such as the Jacobian determinant of a deformation, are highly sensitive to image noise. Conversely, manual Region of Interest (ROI) selection is prone to observer bias and fails to capture the full complexity of the deformation. There is a critical need for automated methods that can simultaneously evaluate fine details and large-scale regional shifts within a statistically sound framework.
Methodology: The Pressure Field Analogy
The proposed framework treats anatomical deformations as a physical system governed by a "pressure field." The methodology rests on three pillars:
- Helmholtz Decomposition: The 3D deformation field (the mapping of one image to another) is decomposed to isolate its irrotational component. This component is specifically responsible for volume changes (expansion and contraction).
- The Pressure Field ($P$): The irrotational component is defined as the gradient of a scalar pressure field. In this model, local maxima in pressure represent "sinks" (areas of contraction or atrophy), while local minima represent "sources" (areas of expansion or growth).
- Probabilistic Scale-Space Analysis: To handle noise and multi-scale effects, the authors apply a Difference of Gaussians (DoG) approach to filter the pressure field at various scales ($\sigma$). A Bayesian framework then calculates the probability of a point being a true extremum.
By evolving a surface to maximize the "flux" of the deformation, the authors define "Critical Regions." This process effectively carves out the most representative "blobs" of change, creating a sparse, simplified, and interpretable map of the underlying biological process.
Key Results and Applications
The framework was validated using two distinct longitudinal datasets: the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and a longitudinal tumor case.
Alzheimer’s Disease Atrophy
In a group-wise analysis of 30 patients, the model successfully localized the most persistent atrophy to the ventricles and hippocampi. Crucially, the scale-space approach allowed for the separation of these large-scale regional changes from more localized cortical thinning, providing a clearer picture of disease progression than standard local metrics.
Tumor Evolution and Decompression
The framework was applied to a patient with a brain tumor treated with Avastin. While standard scans can sometimes mask tumor boundaries during treatment, this method captured the physical reality of the tumor's evolution. It identified both the primary tumor shrinkage (contraction) and the resulting decompression of the surrounding healthy brain tissue (expansion).
Discussion and Future Directions
By focusing on "critical regions," the method discards noisy or irrelevant deformation patterns, leaving a clean summary of anatomical change. The "persistence" of these changes across scales serves as a natural filter for significance; changes that appear across many scales are likely more biologically relevant than those appearing only at a single, fine scale.
While the current framework is fully automated and robust, the authors identify several avenues for future work. Extending the framework from 4D (3D space + 1D scale) to 5D (adding a continuous time dimension) would allow for even more sophisticated longitudinal modeling. Additionally, the authors suggest that the hierarchical "nesting" of these regions across scales could be modeled using graph theory or Morse-Smale complexes to create a structured "tree" of anatomical evolution.
Summary updated