---
title: Dose Volume Constraints in Radiotherapy
url: https://www.emergentmind.com/topics/dose-volume-constraints-dvcs
type: topic
---

# Dose Volume Constraints in Radiotherapy

Dose-volume constraints (DVCs) are treatment-planning criteria that specify a threshold dose and a fraction of a structure that must lie above or below that threshold. In radiotherapy they are evaluated through dose-volume histograms (DVHs) and associated summary indices such as \(V_x\), \(D_y\), mean dose, maximum dose, minimum dose, and percentile dose. Across inverse planning, automated dose prediction, proton robustness, brachytherapy, and radiobiological modeling, DVCs function variously as exact feasibility conditions, conservative surrogates, learned priors, differentiable loss terms, and clinical acceptability criteria [1809.00744][2603.29670][1703.09461].

## 1. Formal definitions and DVH-based metrics

A standard mathematical definition treats a DVC as a percentile constraint on voxel doses. For a structure \(s\) with voxel index set \(\mathcal{V}_s\), the value count function is
\[
v_s(y,b)=\sum_{i \in \mathcal{V}_s} \mathbbm{1}\{y_i \ge b\},
\]
and the \(p\)-th percentile dose is
\[
D_s(p,y)=\max\{b\in\mathbb{R}_+ : v_s(y,b)\ge \phi_s(p)\}, \qquad \phi_s(p):=\frac{p}{100}\,|\mathcal{V}_s|.
\]
An upper DVC \(D_s(p,y)\le u\) is equivalent to \(v_s(y,u)\le \phi_s(p)\), while a lower DVC \(D_s(p,y)\ge \ell\) is equivalent to \(v_s(y,\ell)\ge \phi_s(p)\). In this formulation, DVCs are also called percentile constraints [1809.00744].

Clinical implementations usually express the same idea with DVH-derived metrics. In the head-and-neck setting, the clinical plan evaluation template in the clinical DVH metric loss paper lists \(D_{x\%}\), \(D_{x\,\text{cc}}\), \(D_{\max}\), \(D_{\min}\), \(D_{\mean}\), \(V_{x\%}\), and \(V_{x\,\text{Gy}}\). The paper defines
\[
D_{x\%} = d^{\downarrow}_k, \qquad k = \left\lceil \frac{x}{100}N \right\rceil,
\]
\[
D_{x\,\text{cc}} = d^{\downarrow}_k, \qquad k = \left\lceil \frac{x}{V_{\text{voxel}}} \right\rceil,
\]
\[
D_{\max}=d_1^\downarrow, \qquad D_{\min}=d_N^\downarrow, \qquad D_{\mean}= \frac{1}{N}\sum_{i=1}^N d_i,
\]
and
\[
V_{x\%} = \frac{1}{N}\sum_{i=1}^N H(d_i - T),
\]
with the same form for \(V_{x\,\text{Gy}}\) under an absolute threshold \(T\) [2603.29670].

The metric repertoire is modality-independent. In high-dose-rate prostate brachytherapy, the clinically used summary statistics are \(D_x\), \(V_y\), and \(D_{\max}\). The local protocol cited there uses PTV \(V_{100\%} \ge 90\%\), rectum \(D_{10\%} \le 7.2\) Gy and \(D_{\max} \le 8\) Gy, and urethra \(D_{10\%} \le 10\) Gy and \(D_{\max} \le 10.6\) Gy; a shell structure 2 mm around the prostate is assigned \(D_{\max} \le 8.5\) Gy [1703.09461]. In head-and-neck dose prediction, DVH-based constraint surrogates include OAR Mean[Gy] and \(D_{0.1\text{cc}}\), and PTV \(D_{99\%}\), \(D_{95\%}\), and \(D_{1\%}\) [2111.00077].

## 2. Nonconvexity and direct inverse-planning formulations

The central mathematical difficulty is that clinically relevant DVCs are nonconvex. In the convex optimization treatment, the source of nonconvexity is the indicator in the count function \(v_s(y,b)\): changing one voxel across a threshold changes the count discontinuously, so exact DVC enforcement becomes a subset-selection problem [1809.00744]. The nonconvex IMRT formulation makes this explicit by writing, for an OAR \(j\),
\[
\|(A_j x - d_j^{\mathrm{dv}})_+\|_0 \le \frac{n_j p_j}{100},
\]
where only a limited number of voxels may exceed the threshold dose. The paper notes that this is a cardinality constraint and relates it to the general \(\ell_0\)-constrained problem, which is NP-hard in general [1907.10712].

One route is to encode the DVC directly as a sparsity-constrained split feasibility problem. In the split-feasibility formulation for IMRT/IMPT, the ordinary dose constraints are collected into
\[
C := \{x\in\mathbb{R}^n \mid c \le Ax \le b\}\cap \mathbb{R}_+^n,
\]
while the DVC is imposed in dose space as
\[
Q := \{y\in\mathbb{R}^{m_1}\mid \|(y-b^1)_+\|_0 \le \alpha m_1\}.
\]
The inverse-planning problem then becomes: find \(x^\ast\in C\) such that \(A_1x^\ast\in Q\). The paper describes this as a new type of sparsity constraint that enables the inclusion of a percentage-violation constraint in the model problem and its handling by continuous methods [1702.07925].

Another direct route appears in HDR prostate brachytherapy. There, the mixed-integer model maximizes
\[
\frac{1}{|I_{PTV}|} \sum_{i \in I_{PTV}} v_i,
\]
which is exactly \(V_{100\%}\) for the PTV, subject to OAR fraction constraints
\[
\sum_{i \in I_s} v_i \ge \tau_s |I_s| \qquad \forall s \in \{R,U,S\},
\]
and a dwell time modulation restriction
\[
t_{j_1} \le \gamma t_{j_2}, \qquad \gamma = 2,
\]
for adjacent dwell positions in the same catheter. The simulated-annealing method DOPSA then searches for a dwell-time vector satisfying all DVCs and the DTMR while maximizing PTV coverage [1703.09461].

## 3. Conservative reformulations, two-pass methods, and LP surrogates

Because exact DVCs are nonconvex, several papers replace them with conservative or relaxational surrogates. In the convex framework, the indicator \(g(u)=\mathbbm{1}\{u\ge 0\}\) is upper-bounded by
\[
\hat g_\lambda(u)=(1+\lambda u)_+ = \max(1+\lambda u,0), \qquad \lambda>0,
\]
which yields the convex restriction
\[
v_s(y,b)\le \hat v_s(y,b;\lambda) :=\sum_{i\in\mathcal{V}_s}(1+\lambda(y_i-b))_+.
\]
For an upper DVC \(D_s(p,y)\le u\), the restriction becomes
\[
\sum_{i\in\mathcal{V}_s}(\alpha+(y_i-u))_+ - \alpha\,\phi_s(p)\le 0, \qquad \alpha\ge 0.
\]
This restriction is conservative by construction. To reduce conservatism, the method introduces a two-pass planning algorithm: the first pass solves the convex restricted problem, and the second pass replaces the surrogate with exact voxel bounds on a selected subset so that the dose-volume constraint can be met exactly if the corresponding restriction is feasible on the first pass. A slack-variable variant relaxes lower bounds as \(\ell \mapsto \ell-\delta\) and upper bounds as \(u \mapsto u+\delta\). The implementation is provided in ConRad, a Python-embedded open-source package built on CVXPY with solvers such as SCS and ECOS [1809.00744].

A different surrogate family uses C-VaR-type constraints. The successive LP approach begins from the Romeijn et al. replacement of DVCs with C-VaR-style bounds. For an upper DVC, the surrogate is
\[
\overline{\zeta}_s^\alpha + \frac{1}{\alpha |V_s| }\sum_{i=1}^{|V_s|} (z_{si} - \overline{\zeta}_s^\alpha)^+ \le U_s^\alpha,
\]
and for a lower DVC,
\[
\underline{\zeta}_{s}^{\alpha} - \frac{1}{(1-\alpha) |V_s|} \sum_{i = 1}^{|V_s|} ( \underline{\zeta}_{s}^{\alpha} - z_{si} ) ^{+} \geq L_s^{\alpha}.
\]
The paper emphasizes that the C-VaR feasible region is much narrow compared to the region intended by the DVCs, largely because a small number of outlier voxels can dominate the tail average [1612.01362].

The successive LP refinement enlarges that feasible search region by detecting outliers from the solution of LP problems and removing them from the domain of the C-VaR type constraints. Its key theorem states that if the optimal value \(t^{(k)}\) of the \(k\)-th LP is non-positive, then the corresponding solution satisfies all the DVCs. On the TG119 cases, the method succeeded on three of four cases and typically reached a feasible solution in 2–3 LP iterations, whereas the experiments were capped at 5 iterations; the Romeijn LP was infeasible on all four TG119 cases in preliminary experiments [1612.01362].

## 4. Data-driven, atlas-based, and differentiable representations

Not all modern systems treat DVCs as manually entered hard constraints. The atlas-based voxel dose prediction framework infers a dose-volume prior from similar patients and uses a conditional random field to enforce a learned joint dose prior over ROIs. The central inference problem is
\[
\tilde{d}_{a} = \arg\max_{d_{a,*}} \prod_x P_{*,j}(d_{a,x}\mid F_{*,a,x})\,P(d_{a,*}\mid R(x)),
\]
and the CRF constrains the number of voxels in ROI \(R\) assigned to dose bin \(g\) through
\[
\sum_{x \in R} b_{x,g} \le |R|\,P(d_{a,*,g}\mid R(x)).
\]
In this formulation, DVCs are not manually defined as scalar thresholds; they are inferred as probabilistic dose-volume targets from selected atlases. The method reduced the average MAD in DVHs from 2.57 to 1.91 and showed that atlas selection particularly improved Whole Breast, Prostate, and Rectum cases [1608.04330].

Deep learning models often replace hard DVCs with differentiable surrogates. DeepDoseNet compares MSE, MAE, and MAE+DVH supervision for 3D dose prediction. The DVH score is the mean absolute difference of Mean[Gy], \(D_{0.1\text{cc}}\), \(D_{1\%}\), \(D_{95\%}\), and \(D_{99\%}\) between prediction and reference, and the MAE+DVH model achieved validation \(\bar{S_D}=2.3\) Gy and \(\bar{S_{DVH}}=1.9\) Gy, and test \(\bar{S_D}=2.0\) Gy and \(\bar{S_{DVH}}=1.6\) Gy. It significantly outperformed MAE and MSE for OAR \(D_{0.1\text{cc}}\), particularly for mandible and parotids, and for target \(D_{99\%}\), \(D_{95\%}\), and \(D_{1\%}\) [2111.00077].

A related strategy avoids DVH thresholding directly. The moment-based loss paper approximates DVH behavior through structure-wise dose moments
\[
M_{p} = \left(\frac{1}{|V_s|}\sum_{j\in V_s} d_j^p \right)^{\frac{1}{p}},
\]
and uses
\[
L_{\text{Total}} = L_{MAE} + w_{\text{Moment}}L_{\text{Moment}}.
\]
With \(P=\{1,2,10\}\), the method improved DVH-score by 11% relative to MAE and by 8% relative to MAE+DVH, while improving computational cost by 48% relative to the sigmoid-based DVH loss [2207.03414].

The most direct clinically guided formulation in this group is the clinical DVH metric loss. It optimizes the exact quantities used in plan approval rather than the full DVH curve:
\[
\mathcal{L}_{\mathrm{CDM}} = \sum_{r\in\mathcal{R}}\sum_{k\in\mathcal{M}_r} w_{r,k}\,\left|M^{\mathrm{pred}}_{r,k} - M^{\mathrm{gt}}_{r,k}\right|,
\]
combined with voxel-wise MAE as
\[
\mathcal{L}_{\mathrm{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}} + \lambda_2 \mathcal{L}_{\mathrm{CDM}},
\]
with \(\lambda_1=1\) and \(\lambda_2=0.5\). Using a standard 3D U-Net, MAE + CDM reduced the PTV Score from 1.544 to 0.491, was the only configuration to satisfy both \(V_{95\%}\ge 98\%\) and \(D_{0.03\text{cc}}\le 110\%\) across all test cases, and, with bit-mask encoding, reduced training time by 83\% [2603.29670].

## 5. Robust, proton-specific, and biologically informed extensions

In proton stereotactic radiosurgery and stereotactic body radiotherapy, DVCs may function primarily as institutional pass/fail criteria rather than as a new mathematical object. The GPU-accelerated VPMC study states that its “institutional guidelines on the dose volume constraints (DVCs) for PBSPT-based SRS/SBRT in the treatment of brain cancers” were used, based on the institution’s photon SRS/SBRT protocol. The explicitly reported target criteria are \(V_{100\%}\ge 99\%\), \(D_{\min}\ge 90\%\), \(CI_{100\%}\le 1.5\sim 2.0\), and \(RHI_{100\%}\le 1.3\sim 1.4\), where \(CI_{x\%}\) is the ratio of the volume enclosed by the \(x\%\) iso-dose to the volume of the structure of interest and \(RHI_{x\%}\) is the ratio of the maximum dose within the structure of interest to the \(x\%\) prescription level. Robust optimization modeled range uncertainty by scaling the CT relative-to-water stopping-power ratio by \(\pm 3\%\) and setup uncertainty by shifting the isocenter by 2 mm in all three cardinal directions, giving 13 uncertainty scenarios. The robustly optimized plans using the enhanced VPMC met all DVCs, whereas with the original VPMC without aperture modeling the plans did not always meet the target DVCs [2307.01416].

A stronger extension generalizes DVCs from dose alone to the joint distribution of dose and LET. The dose-LET volume constraint paper defines the dose-LET volume histogram index as
\[
V_{D,\mathrm{LET}}(d,l)=V(D\ge d,\ \mathrm{LET}\ge l),
\]
and treats DLVCs as soft constraints in robust optimization. In this formulation, standard DVHs and LETVHs are replaced by a DLVH surface over the dose-LET plane, and xBD is used as an additional joint metric. For prostate SSPT, DLVCRO significantly improved dose, LET and xBD distributions to protect OARs in nominal and worst-case scenarios while keeping target physical dose distributions comparable [2405.03916].

Biological reinterpretation can also be data-derived. In the causal machine learning analysis of mandible osteoradionecrosis, the treatment variable is a dosimetric factor binarized by a threshold, the covariates are clinical factors, and the outcome is ORN incidence. The paper defines
\[
ATE = E[Y(T = 1) - Y(T = 0)],
\]
and defines the DVC for each dosimetric endpoint as the threshold producing the largest ATE, scanning thresholds in 1% increments and estimating 95% confidence intervals with 1,000 bootstrap iterations. The critical volumes were smaller for PBSPT than for VMAT at every evaluated dose level; for PBSPT they were 35.17 cc at \(V_{40\text{ Gy}[RBE]}\), 32.39 cc at \(V_{50\text{ Gy}[RBE]}\), 17.26 cc at \(V_{60\text{ Gy}[RBE]}\), and 4.00 cc at \(V_{70\text{ Gy}[RBE]}\). Tolerance-curve matching then yielded empirical RBE values of 1.611 at 40 Gy[RBE=1.1], 1.300 at 50 Gy, and 1.132 at 60 Gy, with the paper concluding that RBE exceeded 1.1 in the moderate dose range [2509.21669].

## 6. Fractionation, dose conversion, and cumulative DVH estimation

DVC evaluation becomes more difficult when dose distributions must be compared across fractionation schemes. The voxel-based To2GyConstraints method is designed specifically to make hypofractionated dose distributions interpretable using conventional 2 Gy-per-fraction OAR constraints. The method uses the Linear Quadratic model for doses per fraction \(\le 7.5\) Gy and the Linear Quadratic-Linear model for higher doses, defines a threshold as the mean between the prescribed hypofractionated fraction dose and 2 Gy, and then applies EQD2-like conversion above the threshold and EQDx-like conversion with equivalent-fraction reconstruction below it. For the prostate proof-of-concept of 60 Gy in 20 fractions, the prescription fraction dose is 3 Gy, so the practical threshold is 2.5 Gy/fraction. The paper reports that, for doses above the threshold, To2GyConstraints behaved like EQD2 and yielded clinically realistic dose values, whereas for doses below the threshold it behaved more like BED and avoided the low-dose underestimation seen with standard EQD2 [2604.10227].

The cumulative-DVH problem under anatomy change and day-to-day variation has been treated as an optimal transport problem on time-dose-volume space. The TDV framework defines a Riemannian manifold with coordinates \((T,D,V)\) and metric
\[
ds^2 = dT^2 + \xi^2(T)\, dD^2 + \eta^2(T)\, dV^2.
\]
The method imposes the cumulative monotonicity condition
\[
V(D)\ge V(D+\Delta D), \qquad \Delta D\ge 0,
\]
and partially Ricci-flat conditions so that dose and volume remain metrically equivalent during transport. It uses the \(p\)-Wasserstein distance and derives a transport equation that yields a feasible total DVH. In the prostate IMRT example with rectum and bladder over 20 fractions, the resulting total DVH was compared against the plan and the average of the 20 fractions [2407.19876].

These two directions address different but related DVC problems. The conversion paper translates hypofractionated voxel doses into the language of conventional OAR constraints, whereas the optimal-transport paper estimates a feasible cumulative DVH from time-varying fraction DVHs. This suggests that, beyond static inverse planning, DVC methodology increasingly includes radiobiological state conversion and temporally consistent DVH aggregation [2604.10227][2407.19876].

## 7. Recurring limitations, misconceptions, and open issues

A recurring misconception is that DVCs are always explicit hard constraints in the optimizer. Several papers use them differently. In the VPMC aperture-block study, DVCs are institutional clinical planning requirements used to judge whether a robustly optimized plan is acceptable, and only target DVCs are listed because OAR DVCs were “easily met” for the three selected patients [2307.01416]. In DeepDoseNet, DVCs are represented as evaluation metrics derived from predicted dose distributions rather than as hard constraints [2111.00077]. In the atlas-based framework, they are inferred as probabilistic dose-volume priors rather than manually entered rules [1608.04330].

A second recurring issue is conservatism. The convex restriction of percentile constraints is sufficient but not necessary, so it can exclude feasible exact-DVC solutions [1809.00744]. The C-VaR paper states that Romeijn et al.’s LP feasible region was much narrow compared to the DVCs and that the approach often failed to find a feasible plan even when the DVCs were not so tight [1612.01362]. These observations explain the repeated appearance of two-pass methods, slack variables, outlier removal, polishing, and nonconvex relaxations.

A third recurring issue is that DVH agreement or dose prediction quality is not identical to plan deliverability or plan acceptability. DeepDoseNet explicitly notes that predicted dose distributions are not the same as clinically deliverable plans, that dose prediction is not Pareto optimal planning, and that final plans still require clinical review and possible manual adjustment [2111.00077]. The clinical DVH metric loss paper addresses this mismatch by directly optimizing the exact metrics used for plan approval rather than whole-DVH or voxel-wise losses [2603.29670].

Finally, several papers point to the limits of dose-only DVCs. The DLVH work argues that separately constraining dose and LET can miss the biologically important overlap of the two in SSPT [2405.03916]. The causal ORN study reports smaller critical mandible volumes for PBSPT than for VMAT and empirical RBE above 1.1 in the moderate-dose range, while also noting residual confounding after propensity matching [2509.21669]. The aperture-block proton study notes additional limitations, including use of a non-divergent aperture, omission of numeric OAR DVC thresholds, demonstration in only three representative robust-optimization patients, and the fact that the aperture was based on the largest tumor cross-section [2307.01416]. Together, these studies indicate that DVCs remain indispensable, but their reliable use increasingly depends on how uncertainty, biology, spatial structure, and delivery context are modeled.

Source: https://www.emergentmind.com/topics/dose-volume-constraints-dvcs