Understanding DSC Deconvolution¶

DSC-MRI uses deconvolution to extract perfusion parameters from the measured signal.

The Measurement¶

DSC-MRI tracks a gadolinium bolus through brain tissue using T2*-weighted imaging. The signal drops as contrast passes through:

DSC signal-time curve showing signal drop during gadolinium bolus passage.

From Signal to Concentration¶

T2* Signal Model¶

Gadolinium causes T2* shortening:

\[ S(t) = S_0 \cdot e^{-TE \cdot \Delta R_2^*(t)} \]

Solving for ΔR2*:

\[ \Delta R_2^*(t) = -\frac{1}{TE} \ln\left(\frac{S(t)}{S_0}\right) \]

Concentration Relationship¶

ΔR2* is proportional to contrast concentration:

\[ \Delta R_2^*(t) = r_2^* \cdot C(t) \]

Where r2* is the transverse relaxivity (osipy defaults: 32 s⁻¹mM⁻¹ for tissue, 50 s⁻¹mM⁻¹ for blood at 1.5T).

The Indicator Dilution Problem¶

Tissue Response¶

The tissue concentration C(t) is the convolution of the arterial input function (AIF) with a tissue response:

\[ C(t) = CBF \cdot \left[ C_{AIF}(t) \otimes R(t) \right] \]

Where R(t) is the residue function (fraction of contrast remaining in tissue).

What is the Residue Function?¶

R(t) describes how contrast leaves tissue:

Residue function R(t) decaying from 1 toward 0, with MTT marked.

R(0) = 1: All contrast is present initially
R(∞) = 0: Eventually all contrast leaves
Shape depends on vasculature

The Deconvolution Problem¶

Forward Problem (Easy)¶

Given CBF, AIF, and R(t), calculate C(t):

\[ C(t) = CBF \cdot \int_0^t C_{AIF}(\tau) \cdot R(t-\tau) d\tau \]

Inverse Problem (Hard)¶

Given C(t) and AIF, find CBF and R(t):

\[ R(t) = \frac{1}{CBF} \cdot \text{deconvolution}[C(t), C_{AIF}(t)] \]

This is ill-posed: small noise causes large errors.

Matrix Formulation¶

Discrete Convolution¶

In discrete form, convolution becomes matrix multiplication:

\[ \mathbf{C} = CBF \cdot \mathbf{A} \cdot \mathbf{R} \]

Where:

\[ \mathbf{A} = \begin{pmatrix} C_{AIF}(0) & 0 & 0 & \cdots \\ C_{AIF}(1) & C_{AIF}(0) & 0 & \cdots \\ C_{AIF}(2) & C_{AIF}(1) & C_{AIF}(0) & \cdots \\ \vdots & & & \ddots \end{pmatrix} \cdot \Delta t \]

Solving the System¶

The deconvolution is:

\[ \mathbf{R} = \frac{1}{CBF} \cdot \mathbf{A}^{-1} \cdot \mathbf{C} \]

But A⁻¹ amplifies noise → need regularization.

SVD-Based Deconvolution¶

Why SVD?¶

Singular Value Decomposition provides a stable way to invert A:

\[ \mathbf{A} = \mathbf{U} \cdot \mathbf{S} \cdot \mathbf{V}^T \]

Where S is diagonal with singular values σ₁ > σ₂ > ... > σₙ.

Truncation for Regularization¶

Small singular values amplify noise. Solution: truncate them.

\[ \mathbf{A}^{-1}_{truncated} = \mathbf{V} \cdot \mathbf{S}^{-1}_{truncated} \cdot \mathbf{U}^T \]

Where σᵢ < threshold × σ₁ are set to zero.

Deconvolution Methods¶

sSVD (Standard SVD)¶

Direct SVD of the convolution matrix:

Pros: Simple, widely used
Cons: Sensitive to bolus delay
Threshold: Typically 0.1-0.2

cSVD (Circular SVD)¶

Uses block-circulant matrix (assumes periodic signal):

Pros: Delay-insensitive
Cons: May underestimate CBF
Note: Commonly used in clinical software

oSVD (Oscillation-Index SVD)¶

Selects threshold based on oscillation in R(t):

Pros: Adaptive, noise-tolerant
Cons: More complex
Criterion: Minimize oscillations while preserving CBF

oSVD oscillation index optimization

# oSVD optimizes threshold to minimize oscillation index:
OI = (1/R_max) * sum(|R(i+1) - R(i)|)
# Choose threshold that balances low OI with accurate CBF

Extracting Perfusion Parameters¶

From the Residue Function¶

Once R(t) is estimated:

Parameter	Calculation
CBF	max(R(t)) × scaling factor
CBV	Area under C(t) ÷ Area under AIF
MTT	CBV / CBF (central volume theorem)
Tmax	Time to maximum of R(t)
TTP	Time to minimum of S(t)

CBF Calculation¶

CBF = max[CBF × R(t)] = CBF × max[R(t)]

If R(t) is properly normalized (R(0)=1):

\[ CBF = max(\mathbf{R}) \]

CBV Calculation¶

From conservation of tracer:

\[ CBV = \frac{\int_0^\infty C_{tissue}(t) dt}{\int_0^\infty C_{AIF}(t) dt} = \frac{AUC_{tissue}}{AUC_{AIF}} \]

MTT via Central Volume Theorem¶

The Central Volume Theorem relates:

\[ MTT = \frac{CBV}{CBF} \]

This is a fundamental relationship in indicator dilution theory.

Delay and Dispersion¶

Bolus Delay¶

If tissue is fed by a delayed artery:

Signal arrives later
Standard deconvolution interprets this as low CBF
Use cSVD or delay-corrected methods

Side-by-side comparison of AIF and delayed tissue concentration curves.

Bolus Dispersion¶

As blood flows through vessels, the bolus spreads:

Peak concentration decreases
Duration increases
Affects CBF estimation

Leakage Correction¶

The Problem¶

In tumors with BBB breakdown, contrast leaks into tissue:

T1 effect: Leakage increases signal (opposite to T2* effect)
T2* contamination: Altered concentration curve

Boxerman-Schmainda-Weisskoff (BSW) Correction¶

Models the leakage contribution:

\[ \Delta R_2^*(t) = K_1 \cdot \overline{\Delta R_2^*_{brain}}(t) + K_2 \cdot \int_0^t \overline{\Delta R_2^*_{brain}}(\tau) d\tau \]