📄 medicine

A dual-input DICOM benchmark for perfusion algorithm validation: temporal-separation bias in single-input deconvolution

This study introduces a dual-input DICOM benchmark to demonstrate that applying single-input deconvolution to tissues with temporally separated blood supplies significantly underestimates blood flow due to temporal-separation bias, while blood volume estimates remain accurate.

Original authors: Tomoki Saka

Published 2026-08-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tomoki Saka

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Medical imaging allows doctors to watch blood flow through the body in real time, turning a static picture of an organ into a dynamic map of life. To understand how well an organ is working, specialists often look at three specific numbers: how much blood is flowing through the tissue, how much blood is sitting inside it at any given moment, and how long it takes for that blood to travel through. These measurements are vital for diagnosing diseases in organs like the liver, where blood arrives from two different sources: a direct line from the heart and a slower, filtered route from the digestive system. For decades, computer programs have been designed to calculate these numbers by comparing the tissue's reaction to the blood's arrival. However, these programs were built on a simple assumption: that all the blood comes from a single source and arrives at the same time. When the reality is more complex, with blood arriving from two directions at different speeds, the old tools might be giving doctors a misleading picture.

A researcher at Tokyo Denki University, Tomoki Saka, set out to test exactly how much these standard computer programs struggle when faced with this dual-source reality. Instead of using patients or animal models, which would introduce too many unpredictable variables, the researcher built a perfect, computer-generated simulation. This digital phantom acts like a flawless test subject, where every single detail of the blood flow is known in advance. The researcher created a massive library of 735 different scenarios, each representing a unique combination of blood flow speeds, delays, and volumes. In this simulation, one stream of blood arrives directly, while a second stream is deliberately slowed down and spread out, mimicking the natural delay seen in organs like the liver. The goal was to see what happens when a standard, single-source computer program tries to analyze this dual-source data.

The results revealed a significant blind spot in current medical technology. When the computer program analyzed the dual-source simulations, it performed reasonably well at calculating the total amount of blood sitting inside the tissue. However, it failed dramatically when trying to measure how fast that blood was flowing. In cases where the two blood supplies were separated in time, the program underestimated the total blood flow by roughly 23 to 24 percent. This error grew larger as the second stream of blood became more dominant. The program essentially saw the delayed blood as a separate, slower event and ignored it when calculating the peak speed, reporting a flow rate that was far too low. Even more surprisingly, the researcher found that a more advanced version of the software, which was specifically designed to fix timing errors, offered no help at all. This improved software could handle simple delays, but it could not solve the problem of two distinct streams of blood arriving at different times and merging in a way that confused the calculation.

The study also clarified why this happens. The computer program works by looking for the highest peak in the blood's arrival curve and assuming that peak represents the total speed of the flow. In a single-source system, this works perfectly. But when two streams arrive at different times, the second stream arrives later and is spread out, creating a broad, low hill rather than a sharp peak. The computer, looking for that sharp peak, only sees the first stream and misses the contribution of the second. The total volume of blood remains accurate because the computer correctly adds up the entire area under the curve, but the speed calculation collapses because it relies on a single, sharp moment that no longer exists. This finding suggests that for organs with dual blood supplies, the standard method of measuring flow speed is fundamentally flawed when the two supplies are not perfectly synchronized.

The researcher confirmed that this problem is not a glitch in the software's settings but a fundamental limitation of the method itself. The error persisted regardless of how the data was sampled or how the computer smoothed out the numbers. The study did not offer a new solution or a new program to fix the issue; instead, it provided a rigorous benchmark and a clear warning. It showed that while current tools can still reliably measure blood volume in these complex organs, the flow speed numbers they produce are likely to be significantly too low. For doctors relying on these numbers to make critical decisions about liver health, the study indicates that the flow measurements may need to be interpreted with extreme caution, or that entirely new methods must be developed to account for the two separate paths blood takes through the body.

Technical Summary: A Dual-Input DICOM Benchmark for Perfusion Algorithm Validation

Problem Statement
Perfusion imaging algorithms, typically used in CT or MR, estimate blood flow (BF), blood volume (BV), and mean transit time (MTT) by deconvolving a tissue concentration curve with an arterial input function (AIF). Established validation benchmarks, such as the widely used Kudo 2013 digital phantom, rely on a single-input model where tracer delay is simulated by shifting the entire tissue curve. While mathematically exact for single-input linear time-invariant systems, this "frame-shift" construction fails to represent organs with dual blood supplies (e.g., the liver with hepatic artery and portal vein, or the lung) where distinct supply routes arrive with different delays and dispersion characteristics.

When single-input deconvolution is applied to such dual-supply tissues, the resulting "effective" impulse response is a superposition of the arterial and second-route contributions. If these contributions are temporally separated, the peak of the effective response no longer represents the total physical flow, leading to significant model-mismatch bias. Existing delay-insensitive algorithms (like block-circulant SVD) address shifts between a single input and tissue but do not resolve the bias introduced by the temporal separation of two distinct inputs.

Methodology
The author constructed an independently generated dual-input DICOM benchmark following the layout of the Kudo 2013 phantom but with closed-form ground truth for a dual-supply system.

Phantom Generation: The phantom consists of 735 parameter combinations arranged in a tiled structure. It features a fixed arterial component and a variable second route (simulating portal-venous supply). The second route is modeled as an area-conserving dispersion of the arterial input (using an exponential kernel with time constant $\tau$ ) plus an additional onset delay ( $\Delta_1$ ).
Synthesis: Tissue curves were synthesized on a fine temporal grid (0.05 s) to accurately capture sub-frame delays and dispersion, then decimated to the observation grid (2.0 s). The curves were embedded as uniform tile values in DICOM CT Image Storage objects without added noise or physical acquisition artifacts.
Ground Truth: True values for BF, BV, and MTT were defined based on the central volume principle applied to the composite system.
Evaluation: Two standard single-input deconvolution algorithms were applied using only the arterial input: standard SVD (sSVD) and block-circulant SVD (bSVD).
Analysis Strategy: To isolate the specific bias caused by temporal separation, the study compared each dual-input condition against a "synchronous control" (where $\tau=0$ and $\Delta_1=0$ ) with identical residue shapes, volumes, and transit times. This allowed the calculation of $\Delta E_{sep}$ , the error attributable solely to the temporal arrangement of the two supplies, independent of general SVD regularization bias.

Key Results

Blood Volume (BV): BV estimates tracked ground truth closely (regression slope $\approx$ 0.98 for sSVD, 0.99 for bSVD). This confirms that the zeroth moment of the effective response (total area) remains conserved despite temporal separation, provided the second route is area-conserving.
Blood Flow (BF): BF estimates showed severe underestimation and poor correlation with ground truth (regression slope $\approx$ 0.16 for sSVD, 0.13 for bSVD). The peak of the effective response failed to capture the sum of the two physical flows when they were temporally separated.
Temporal Separation Bias: Isolating the effect of temporal separation revealed that it alone reduced estimated BF by approximately 23.2 percentage points for sSVD and 24.2 points for bSVD. This error scaled linearly with the flow fraction of the second route ( $\phi_1$ ), with a correlation of $r = -0.86$ for sSVD.
Algorithm Robustness: The bias persisted across SVD truncation thresholds (5–20%), sampling intervals (1–3 s), and observation windows (58–120 s). Crucially, bSVD, designed to be delay-insensitive, offered no protection against this specific bias and was marginally worse than sSVD.
Mechanism: The study demonstrates that the failure is not due to a shift between the input and tissue curve (which bSVD corrects) but rather the inability of a scalar peak estimator to sum two temporally separated flow components.

Significance and Claims
The paper claims to provide the first DICOM benchmark specifically designed to validate perfusion algorithms against a dual-input ground truth with known closed-form solutions. Its primary contributions are:

Quantification of Bias: It establishes that for temporally separated supplies, the peak of the arterial-referenced effective response does not represent total flow, while the zeroth moment (BV) remains interpretable.
Limitation of Current Methods: It demonstrates that "delay-insensitive" deconvolution methods (like bSVD) do not solve the problem of dual-input temporal separation, as the issue is not a simple time shift but a structural mismatch in how the effective response is formed.
Validation Tool: The released benchmark allows developers to characterize how much their reported flow values drop as a function of the second-route flow fraction, providing a metric for algorithm performance in dual-supply organs without requiring patient data.

The author remains modest regarding clinical extrapolation, noting that the study is entirely computational, lacks noise modeling, and tests flow fractions up to 0.63 (below some normal liver portal fractions). The benchmark is presented as a tool for algorithm validation and the development of future multi-input estimators, rather than a solution for current clinical software interoperability.

Technical Summary: A Dual-Input DICOM Benchmark for Perfusion Algorithm Validation

More like this