TL;DR
This study uses advanced diffusion MRI models on ex vivo breast cancer tissue to reveal that simple ADC and DTI models are insufficient for characterizing tissue microstructure. More complex models incorporating restriction and anisotropy better explain the data, suggesting that cell size and extracellular matrix organization are crucial for accurate tumor characterization.
What Problem Does This Paper Address?
Current clinical diffusion MRI for breast cancer primarily relies on the Apparent Diffusion Coefficient (ADC) and Diffusion Tensor Imaging (DTI). While these methods can differentiate between benign and malignant lesions, they often show significant overlap in values, limiting their diagnostic accuracy. This is partly because these models are too simplistic to capture the complex microstructure of breast tissue, which includes variations in cell density, immune cell infiltration, and the extracellular matrix. The paper aims to address this by exploring more sophisticated diffusion models that can better represent these microstructural features and improve tumor characterization.
Methods
The researchers employed a comprehensive diffusion MRI protocol on formalin-fixed breast cancer samples. This involved acquiring data across a wide range of gradient strengths, durations, separations, and orientations. They then fitted a variety of one- and two-compartment diffusion models to this rich dataset. These models included:
- Ball (ADC): Represents free, unrestricted diffusion.
- Tensor (DTI): Accounts for anisotropic diffusion.
- Zeppelin: A cylindrical tensor model.
- Sphere: Models diffusion restricted by an impermeable membrane.
- Two-compartment models (e.g., Ball-Ball, Zeppelin-Sphere, Tensor-Sphere): These combine different compartment types to represent extracellular and intracellular spaces, with parameters like extracellular and intracellular volume fractions ($f_e$, $f_i$), diffusion coefficients ($D_1, D_2, D_3$), and restriction radius ($R$).
Model selection was performed using the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC). Parameter distributions were analyzed using Markov Chain Monte Carlo (MCMC) methods. The MRI data was meticulously registered with high-resolution histology (H&E staining) to correlate diffusion parameters with microstructural features.
Datasets
The study utilized seven ex vivo breast tissue samples containing invasive breast cancers. These samples included different grades and subtypes (grade 1 ductal/no special type (NST), grade 3 mucinous, and grade 3 NST). The samples were obtained from six patients through the King's Health Partners Cancer Biobank. No new benchmarks were introduced; the focus was on applying advanced modeling to existing tissue samples.
Key Insights and Ideas
- Simple models fall short: Conventional ADC and DTI models were insufficient to explain the diffusion data in most cancerous regions. This highlights a significant limitation in current clinical practice.
- Restriction and anisotropy are key: More complex models incorporating restricted diffusion (like the Sphere model) and anisotropy (like the Zeppelin and Tensor models) were consistently better at explaining the diffusion signal, particularly in cancerous regions.
- Cell size matters more than just density: While ADC often correlated with cellularity, parameters from more complex models (specifically the radius $R$ from the Sphere compartment) suggested that individual cell size and extracellular matrix organization play a significant role in diffusion signal, offering a more nuanced view than just cell density.
- Stromal orientation is detectable: Anisotropy patterns in diffusion MRI corresponded with the orientation of the fibrous stroma in histology, suggesting that diffusion MRI can probe extracellular matrix structure. However, these coherent regions were small (around 1 mm), requiring high resolution for detection.
Key Results
- Model selection: In most voxels, anisotropic two-compartment models (Zeppelin-Sphere and Tensor-Sphere) were the best fit, indicating the presence of both restriction and anisotropy. Unrestricted models (Zeppelin-Ball, Tensor-Ball) were better in some regions, particularly in grade 3 mucinous carcinoma.
- ADC vs. complex models: While ADC maps generally related to cellularity, parameters from the Zeppelin-Sphere model (like $f_1$ and $R$) provided additional information that better explained variations in diffusion signal, especially in regions where ADC did not perfectly correlate with cellularity.
- Anisotropy coherence: Regions of coherent anisotropy were small (approx. 1 mm) and corresponded to stromal orientation patterns. This coherence weakened significantly at lower resolutions (e.g., 2 mm), potentially explaining inconsistencies in previous clinical findings.
- Diffusion coefficients: Diffusion coefficients were lower than those observed in vitro, likely due to fixation and lower scan temperatures, but the relative signal fractions remained similar, suggesting model ranking is preserved.
Limitations and Future Work
- Sample size: The study used a small number of samples, limiting the generalizability of findings across all breast cancer types.
- Fixation and temperature: Formalin fixation and the lower scan temperature (18.4°C) likely affected diffusion coefficients compared to in vivo conditions, though model ranking was preserved.
- Single cell size parameter: Models assumed a single average cell size, which is a simplification. Future work could explore models incorporating cell size distributions or anisotropic restriction.
- Exchange effects: The models assumed no water exchange between compartments, which might not hold true for longer diffusion times.
- Clinical translation: The study used gradient strengths larger than typically available clinically. Future work should investigate how these findings translate to more limited clinical scan protocols and explore the design of clinical protocols that can capture information about the observed radius range (6-9 µm).
- Sub-voxel anisotropy: Detecting small regions of anisotropy may require techniques sensitive to sub-voxel anisotropy or reduced field-of-view sequences.
Summary updated