Mild photothermal therapy heats a tumor to about 42–45 °C, not the 50 °C and above used for ablation. At that temperature heat does not kill cells on its own. Its job is to weaken them: it disables heat-shock proteins (HSP70, a chaperone that helps stressed cells survive) and makes cells more open to a co-delivered drug or catalyst. This works only when the tumor is heated evenly, healthy tissue stays cool, and heating happens while the cells' defenses are down. Current practice meets none of these requirements well. A recent nanoprobe study fires a single broad laser beam during a fixed 10–12 hour window after injection, a window identified offline from photoacoustic imaging. Photoacoustic imaging reads out light-induced sound, and in that study a signal ratio tracks when HSP70 is suppressed (Visualization of HSP70-regulated mild-photothermal therapy). The one closed-form method for shaping deposited power across many sources is a single central convex solver that assumes static tissue (Robust Optimal Power Distribution for Hyperthermia Cancer Treatment).
We propose treating each heat source as an independent learner. The sources might be interstitial laser fibers or separately driven transducers. Each one sets its own power using only its nearby thermometers and a local readout of how sensitized the tissue is. The idea rests on one physical fact. Heat from a source spreads through perfused tissue and fades over a characteristic distance, so each source mostly affects its own neighborhood. When every source maximizes a reward computed from its neighborhood's sensors, the sources are approximately playing a potential game: a game in which each player's selfish improvement also improves one shared objective. We can bound how far the game is from an exact potential game using the tissue's heat-diffusion length. Existing theory for independent learners in such games then says the sources can reach a coordinated heating pattern with no central computer. The benefit we expect is robustness when a source fails, perfusion shifts, or different tumor regions become sensitized at different times.
Core Claim
The claim has three parts. First, multi-emitter mild hyperthermia with incoherent (non-interfering) sources is an ε-potential game, and ε shrinks with the ratio of a source's sensing radius to the thermal diffusion length. Second, independent entropy-regularized natural policy gradient learners, using only local feedback, match a centralized robust optimizer on nominal tissue. Third, those learners degrade less than the centralized optimizer under source dropout, perfusion change, and region-by-region shifts in the sensitization window. The claim is scoped deliberately. For a phase-coherent ultrasound array, every element's contribution interferes at every point, so the locality argument fails and a centralized solver is the natural tool. We include that setting as a regime boundary, not as a target to beat.
Mechanism
Model the tissue with the standard heat equation plus a blood-perfusion heat sink. At quasi-steady state, the temperature rise is linear in the vector of source powers: $\Delta T = G p$. Here $G_{ki}$ is the rise at sensor $k$ per unit power from source $i$. In a diffusion equation with a linear sink, the Green's function decays roughly as $\exp(-d/\ell)$ with distance $d$, where $\ell=\sqrt{\kappa/(w_b c_b)}$. In that expression, $\kappa$ is tissue conductivity, $w_b$ is perfusion rate and $c_b$ is blood heat capacity.
The shared objective is
$$\Phi(p)=\sum_{k\in \text{tumor}} s_k\, \phi(\Delta T_k) \;-\; \lambda \sum_{k\in\text{healthy}} \big[\Delta T_k-\tau\big]_+ ,$$
where each symbol is defined as follows:
- $s_k\in[0,1]$ is the local sensitization level, simulated as a photoacoustic-ratio proxy for HSP70 depletion.
- $\phi$ is a saturating reward for reaching the mild-hyperthermia band.
- $\tau$ is the healthy-tissue tolerance.
- $\lambda$ is a penalty weight.
Source $i$ maximizes $\Phi_i$, which is the same sum restricted to sensors within radius $r$ of source $i$. The mismatch between source $i$'s local reward and its true marginal effect on $\Phi$ comes only from sensors outside that radius. So the potential-game error satisfies
$$\epsilon \;\lesssim\; C\, \max_i \sum_{k:\, d_{ik}>r} G_{ki}\;\sim\; e^{-r/\ell},$$
where $C$ bounds the slopes of $\phi$ and the penalty. This yields a falsifiable prediction that is not obvious in advance: higher perfusion shortens $\ell$ and should make the problem more decentralizable, not less.
We borrow the learner from Independent Natural Policy Gradient Methods for Potential Games. That paper shows that independent softmax natural policy gradient with entropy regularization converges to a regularized equilibrium at a rate that does not depend on action-set size and grows like $O(\sqrt{N})$ in the number of agents. It assumes exact utility oracles. We replace the oracle with a local critic trained on sensor data. Two things differ from the theory's setting:
- Delayed cost. Heat keeps spreading after the power is cut, so an action that looks safe now can overheat healthy tissue a minute later. The local critic therefore predicts discounted future healthy-tissue excess, not the current excess.
- Multiple timescales. Heat evolves over seconds to minutes, while sensitization changes over hours. We treat $s_k$ as a slow context input to the policy rather than as state the critic must track.
Two convergence analyses disagree about which component should adapt faster. The two-timescale analysis of Independent and Decentralized Learning in Markov Potential Games uses a fast critic. The optimistic-gradient analysis of Last-iterate Convergence of Decentralized OGDA uses a slow critic. We test both orderings empirically instead of assuming either.
Relation to Prior Work
The two closest works each supply half of what we need, and neither uses learning in the other's domain.
- Centralized power shaping. Robust Optimal Power Distribution for Hyperthermia solves a worst-case power-shaping problem for 51 coherent transducers as one semidefinite program. It needs global knowledge of an ellipsoidal uncertainty set and never adapts during treatment. We keep its geometry and metrics but replace the single solver with local learners and add time-varying perfusion and sensitization.
- Adaptive timing from imaging. Visualization of HSP70-regulated mild-photothermal therapy motivates using the photoacoustic ratio as a timing signal, but uses it only as an offline gate. We make it an online, spatially resolved input.
- Why heat and chemistry together. In CNSI-Fe for triple-negative breast cancer, 78% tumor growth inhibition came from the iron complex plus near-infrared light, versus 64% for heat alone and 28% for the complex alone. This is why the objective weights heat by chemical sensitization rather than rewarding temperature alone.
- Safe multi-agent learning. Multi-Agent Constrained Policy Optimisation supplies a safe-learning comparison point. It trains with a centralized critic and sequential updates, and it has only been tested on robots whose constraint violations are instantaneous collisions. Our constraint is diffusive and lagged.
Research Program
We will build a 2D simulator of the heat equation with perfusion. We start from the breast geometry of Benchmark E: a 10 cm semicircle containing a 16 mm tumor, monitored at 174 healthy-tissue and 13 tumor control points. We then add three things:
- Sources. Between 8 and 32 incoherent sources, placed on the surface and interstitially.
- Perturbations. Perfusion fields that vary over time, and random source dropout.
- Sensitization. A field $s_k(t)$ whose peak window is centered near the reported 10–12 hour window, with shifts of ±4 h drawn per region.
Each source chooses among 20 discrete power levels, matching the action-set size of Benchmark D. The analysis has three parts: estimating $G$ and $\epsilon$ empirically as functions of $r$ and $\ell$, measuring convergence of independent learners against that bound, and running the robustness comparisons described below.
Baselines
The naive reference is a single uniform beam at fixed fluence, fired during the fixed 10–12 hour window. This mirrors the published protocol of Visualization of HSP70-regulated mild-photothermal therapy.
The strongest existing method is the robust semidefinite program of Robust Optimal Power Distribution for Hyperthermia. Under its perturbed model it reports average tumor power of −27.17 dB and healthy-tissue leakage of −17.43 dB. The nominal design under the same perturbation reports −36.40 dB and −11.69 dB. We will re-implement it on our geometry and give it the true nominal $G$ and a correctly sized uncertainty set. We will also run a version that re-solves every control step with perfect state, which is a deliberately generous central controller. Beating it means matching its tumor coverage on nominal tissue within 1 dB, and achieving lower healthy leakage and higher sensitized-dose coverage under dropout and perfusion shift. The published numbers come from a coherent array, so on our incoherent sources we compare only against our own re-run.
The strongest learning baseline is MAPPO-Lagrangian, run from its official code with a centralized critic that sees all sensors. It gets the same environment steps and the same hyperparameter-search budget as our method. The file reports no numerical costs for it (no reported result found on Lacuna for MAPPO-Lagrangian cost values), only that it achieved near-zero cost where unconstrained baselines failed. Beating it means comparable reward and cost with no global information, and better retention when sources drop out at test time.
The simplest decentralized baseline is one PID temperature clamp per source, targeting 43 °C and gated by the local photoacoustic ratio. It is the cheapest competitor that already uses both local feedback and adaptive timing.
Four ablations isolate the key ingredients:
- Global reward. Independent learners given the global $\Phi$, an oracle upper bound on what locality costs.
- No sensitization. $s_k\equiv 1$, removing the chemistry weighting.
- No entropy. Entropy regularization removed.
- Radius sweep. Neighborhood radius $r$ varied to trace the $\epsilon$ curve.
Why not just…?
Why not just solve the centralized robust program? When the tissue model is known and stationary, it is likely hard to beat, and we expect to tie there. Its weaknesses are structural. It needs a known uncertainty set and a live central link to every source, and it models neither lagged heating nor spatially varying sensitization. The dropout and window-shift experiments measure directly whether those weaknesses matter. The re-solving central variant tells us whether any gap comes from decentralization itself or only from re-planning. If that variant wins everywhere, we will say so.
Why not just use centralized-critic multi-agent learning such as MAPPO-Lagrangian? In simulation, a global state is available, so this is a real threat during training. Our case rests on deployment. Sensors fail, sources get added or removed, and the policy must keep working without the global view it was trained on. The held-out dropout test settles whether training with a centralized critic makes the learned policies fragile.
Why not just put an independent PID clamp on each source? This is the most serious simple alternative, because it is already decentralized. Its expected failure is overlap. Neighboring clamps each push power into the tissue between them, overheating healthy tissue there and leaving cold spots at the tumor edges, because no clamp accounts for its neighbors' heat. If the gated PID matches our learners on sensitized coverage and healthy excess, then learning adds nothing in this regime. That result would redirect the work toward the coherent-array setting.
Why not just train one single-agent reinforcement-learning controller over all sources? Its joint action space has $20^N$ combinations, and it is centralized, so it inherits the central solver's dependence on a live link to every source. We will run it with 8 sources as a scaling reference, not as a main competitor.
Decisive Checks
Three results decide whether the core claim holds:
- The empirical potential-game error should fall exponentially in $r/\ell$. The gap between our learners and the global-reward oracle should track that error. Raising perfusion should shrink the gap.
- When 25% of sources are dropped at test time, local learners should lose less sensitized-dose coverage than both the re-solving central program and MAPPO-Lagrangian.
- With the sensitization window shifted per region, local learners should beat the fixed-window beam and the gated PID on sensitized coverage while keeping healthy tissue below τ.
If check 1 fails, the locality argument is wrong. If checks 2 and 3 fail against the gated PID, the proposal reduces to "clamp each source locally," and we will report that.
References
- Visualization of HSP70-regulated mild-photothermal therapy for synergistic tumor treatment: a precise space-time mild-temperature photothermal ablation strategy
- Robust Optimal Power Distribution for Hyperthermia Cancer Treatment
- Independent Natural Policy Gradient Methods for Potential Games: Finite-time Global Convergence with Entropy Regularization
- Independent and Decentralized Learning in Markov Potential Games
- Last-iterate Convergence of Decentralized Optimistic Gradient Descent/Ascent in Infinite-horizon Competitive Markov Games
- Photothermal therapeutic effects and biosafety of a carbon nanoparticles–Fe(<scp>ii</scp>) complex for triple-negative breast cancer
- Multi-Agent Constrained Policy Optimisation