Diffusion weighted MRI (DWI) is a valuable adjunct to T₂-weighted imaging and dynamic contrast-enhanced MRI to enhance the accuracy of breast cancer diagnosis and staging. Differences in the apparent diffusion coefficient (ADC) of benign and malignant breast lesions have been reported from clinical studies ^1,2, however, the biophysical basis of reduced ADC in cancer tissue remains poorly understood. The most diagnostically useful models are likely to be those based on tissue microstructure. Recent microimaging study of breast tissue ex vivo demonstrated that epithelial cell layers in gland had lower diffusivity than their adjacent supporting stroma ³. Motivated by this finding, this study investigated multi-component restricted and unrestricted models of diffusion in formalin-fixed glandular breast tissue.

DWI. A fixed breast tissue specimen was washed in saline +0.2% v/v Magnevist. Imaging was performed at 22℃ in a Bruker Biospec 9.4T scanner with method similar to ⁴. A pulsed gradient spin echo sequence was applied with: Δ = 10, 20, 40, 80 ms; δ = 5, 10ms; nominal b = 100, 311, 603, 965, 1391, 1873, 2411, 3000 s/mm²; voxel size 1.0×0.8×1.0 mm³; 3 orthogonal gradient directions; TE = 18, 28, 33, 48, 53, 88, 93 ms; TR = 2000ms; SNR_b=0 ~142; FOV 50×30 mm; slice thickness = 1.0 mm; and matrix 50×30. A separate 6-direction DTI acquisition was performed with the same Δ, δ and TE as the 3-direction measurements and nominal b = 965, 1873 s/mm². The data were normalized to reduce T₂ dependence. Effective b-values were used for model fitting.

Diffusion Models. Breast tissue was modeled with one to three components representing non-exchanging spin pools. For each component, there were five candidate models ⁵: 1) a ‘Tensor’ which is a conventional DTI model; 2) a ‘Zeppelin’ that describes a cylindrically symmetric tensor; 3) a ‘Ball’ that is an isotropic tensor; 4) a ‘Sphere’ describing diffusion inside an impermeable sphere of radius R; and 5) a ‘Stick’ assuming diffusion within an idealized cylinder of zero radius. Diffusivities were constrained to be within biologically plausible limits so that 0<D<2.1 μm²/ms. For the ‘Sphere’ model, we constrained the radius to be 0.1<R<20 μm. Eleven models were fitted to each voxel in 68 regions of interest (ROIs) delineating gland lobule tissue (Fig 1A). Note that the ROIs do not include fat and interlobular stroma.

Model fitting and ranking. Each model was fitted to the combined DTI and 3-direction DWI data using the Levenberg-Marquardt minimization algorithm in the open source Camino toolkit ⁶. Akaike Information Criterion (AIC) was calculated to provide an objective comparison of the information content of different models. Lower AIC indicates higher model information content and predicts superior model performance with unseen data.

Fig. 1B presents the overall AIC ranking of the 11 models over all voxels from all gland ROIs (102 voxels total). The models that included the restricted sphere component ranked highest, while the unrestricted Ball (ADC), Bi-ball (biexponential) and DTI models ranked lowest. Fig 2 illustrates the fit of four of the models to a representative voxel with the corresponding AIC ranking position. Table 1 summarizes values of the sphere radius R from the four restricted models.In glandular breast tissue lobules, DWI performed over multiple diffusion times showed distinct evidence of restricted water diffusion. In many voxels the models comprising both a restricted and an anisotropic component ranked higher than the isotropic restricted Ball-sphere model, indicating the presence of significant diffusion anisotropy in one spin pool at this voxel size (~1mm³). Breast gland lobules (the ROIs in this study) are comprised of low diffusivity epithelium and higher diffusivity intralobular fibrous stroma ³. The superior performance of multi-component restricted and anisotropic models we found is consistent with the presence of distinct diffusion microenvironments in breast gland lobules, and the possibility that the restricted and anisotropic spin pools correspond to epithelium and intralobular stroma respectively requires further investigation. This possibility is supported by the mean sphere radius of ~13μm which is similar to the epithelial cell size. Development of clinical DWI methods sensitive to changes in these environments (for example ⁷) may be the key to increased accuracy of breast cancer detection.