Abstract
This paper presents a hierarchical registration framework for 3D breast MR images to correct for non-rigid motion during dynamic contrast-enhanced (DCE) studies. By combining global affine transformations with local B-spline free-form deformations (FFD) and utilizing Normalized Mutual Information (NMI) as a similarity metric, the authors demonstrate a significant reduction in motion artifacts. This approach enhances the reliability of tumor visualization in subtraction images, moving beyond the limitations of standard rigid or affine registration techniques.
The Challenge of Soft Tissue Motion in DCE-MRI
In clinical breast cancer screening, dynamic contrast-enhanced MRI is a primary tool for identifying suspicious lesions. The technique relies on "subtraction imaging," where a pre-contrast image is subtracted from a post-contrast image to highlight areas of contrast agent uptake. Ideally, this process isolates the tumor; however, patient motion between scans—such as breathing or slight shifts in position—introduces significant errors.
Because the breast is composed of soft tissue, it undergoes complex, non-rigid deformations rather than simple linear shifts. Standard rigid registration (translation and rotation) or affine registration (scaling and shearing) cannot account for these local warps. The resulting "motion artifacts" appear as bright or dark halos in subtraction images, which can either mimic tumors (false positives) or obscure actual pathology.
A Hierarchical Transformation Model
To address these deformations, the authors proposed a two-stage hierarchical model:
- Global Motion (Affine): The algorithm first performs a broad alignment using an affine transformation with 12 degrees of freedom. This accounts for general rotation, translation, and scaling of the breast within the scanner.
- Local Motion (Free-Form Deformation): To handle "squishy" tissue warping, the model employs a Free-Form Deformation (FFD) based on B-splines. A 3D grid of control points is overlaid on the image; as these points move, the underlying image volume warps smoothly and realistically.
Normalized Mutual Information (NMI)
A critical innovation in this framework is the use of Normalized Mutual Information (NMI) as the similarity measure. Traditional metrics like Sum of Squared Differences (SSD) fail when image intensities change—which is the very goal of contrast-enhanced imaging. NMI, however, is a statistical measure that identifies structural alignment regardless of global brightness changes, making it exceptionally robust for dynamic studies.
Experimental Results and Clinical Impact
The algorithm was evaluated using both volunteer studies (with intentional simulated motion) and clinical patient data. The results demonstrated a clear performance hierarchy: Non-rigid > Affine > Rigid > No registration.
Quantitatively, the non-rigid model achieved a correlation coefficient (CC) of 0.9899 in volunteer studies, a marked improvement over the 0.9648 achieved by rigid registration. Visually, the impact is most apparent in the subtraction images, where the "ghosting" artifacts are nearly eliminated, leaving only the true contrast uptake visible.
Conclusion
By incorporating B-spline FFDs and NMI, this framework provides a robust solution for the non-rigid registration of breast MRI. The method is purely voxel-based and does not require complex biomechanical modeling of individual patient tissue. While the computational time (approximately 15–30 minutes per study in 1999) was a consideration for clinical throughput, the significant gain in diagnostic clarity established a new standard for motion correction in breast imaging.
Summary updated