Closing the loop — MRI watches the temperature and steers the treatment automatically
Before any energy is delivered to tissue, the MR scanner acquires a baseline phase image of the treatment volume. This single acquisition becomes the zero-point against which every subsequent phase measurement is compared — the entire closed-loop system depends on this reference being clean, motion-free, and precisely registered to the anatomy that will be heated.
MR thermometry does not measure absolute temperature directly — it measures a change in proton resonance frequency relative to a reference state. That reference state is the baseline phase image acquired at normal body temperature, before the hyperthermia applicator is switched on.
Every voxel in the imaging volume has its own baseline phase value, determined by local field inhomogeneities, coil sensitivity, and tissue composition. Because the proton resonance frequency shift (PRFS) method is inherently a differential measurement, any error baked into the baseline — patient motion between acquisitions, respiratory drift, or scanner frequency drift — propagates directly into every temperature map computed afterward.
Clinically, multiple baseline images are often averaged to suppress thermal and system noise, and a low-pass filtered field map is sometimes subtracted to correct for slow B0 drift unrelated to heating.
Because PRFS thermometry is a subtraction method, its accuracy is fundamentally limited by how well the baseline phase image is registered to every later frame — sub-voxel motion between baseline and treatment can introduce several degrees of artificial temperature error.
The baseline acquisition also defines the imaging geometry that will be used throughout the session: which anatomical slice or 3D volume will be tracked, at what in-plane resolution, and how the target tumor or treatment zone is registered relative to the applicator array.
Radiation and hyperthermia planning typically fuses this baseline MR geometry with a pre-treatment planning CT or MR, so the target temperature pattern (commonly a disk or ellipsoid covering the tumor with a small safety margin) can be drawn directly onto the same coordinate frame that the live thermometry will later populate with measured temperatures.
This upfront geometric alignment is what allows the closed-loop controller, several steps downstream, to compare "where the heat is" against "where the heat should be" on a voxel-by-voxel basis without any ambiguity about anatomical correspondence.
Once heating begins, the scanner switches into a rapid-repetition gradient-echo acquisition loop, collecting a new phase image every few seconds. Each new phase frame is subtracted from the baseline to yield a phase difference map — and that phase difference is the raw signal from which temperature change is computed.
The PRFS method exploits a subtle but remarkably linear physical effect: the resonance frequency of water protons is sensitive to local hydrogen bonding. As tissue temperature rises, hydrogen bonds between water molecules stretch and weaken, which increases the electron shielding around each hydrogen nucleus. More shielding means a slightly lower resonance frequency for a given magnetic field — a chemical shift that moves linearly with temperature over the physiological hyperthermia range (up to ~45–50°C).
This coefficient, α, is remarkably consistent across most soft tissues (because it is dominated by tissue water content) at approximately -0.01 ppm/°C, making PRFS one of the few MR thermometry methods that translates well from phantom calibration to in vivo use without per-tissue recalibration. Fat, notably, does not show this effect, which is why fat-suppressed or fat-referenced sequences are used to avoid corrupting the phase map near fatty tissue boundaries.
The measured phase difference Δφ between a heated frame and baseline converts to temperature change ΔT via:
ΔT = Δφ / (γ · α · B0 · TE)
where γ is the gyromagnetic ratio, B0 the main field strength, and TE the echo time — meaning every scanner parameter chosen at setup directly scales how sensitive the measurement is to a given amount of heating.
The PRFS temperature coefficient of roughly -0.01 ppm/°C is one of the most reproducible constants in medical imaging physics — it is what allows MR-guided hyperthermia and MRgFUS systems to report temperature to within about ±1°C without needing an implanted probe.
Designing the real-time acquisition is an exercise in balancing three competing demands simultaneously:
• Spatial resolution: finer voxels resolve sharper hot-spot boundaries and reduce partial-volume averaging at tumor margins, but require more k-space lines per frame, directly slowing acquisition.
• Temporal update rate: the closed-loop controller can only correct deviations as fast as new temperature information arrives. Faster updates (every 1–3 s) let the controller react before a hot spot overshoots, but force compromises in resolution or signal-to-noise ratio (SNR) per frame.
• Motion artifact sensitivity: because PRFS is a phase-subtraction technique, any physiological motion (respiration, swallowing, bowel peristalsis, or even patient repositioning) that is not identical between baseline and live frames introduces phase errors that masquerade as temperature change. Abdominal and thoracic targets often require respiratory-gated or navigator-corrected acquisitions, at the cost of update-rate consistency.
In practice, clinical and academic MR-hyperthermia platforms settle on 2–5 second update intervals with in-plane resolution around 2–3 mm, using parallel imaging acceleration (SENSE/GRAPPA) and multi-echo acquisitions to squeeze more temperature precision out of each frame without sacrificing speed.
The raw phase-difference data is only useful once it is converted, voxel by voxel, into an absolute temperature map overlaid on the anatomical image. This is the moment the abstract MR signal becomes a clinically interpretable picture — revealing exactly where the tissue is heating as planned, where hot spots are forming, and where cold, under-treated regions persist.
Each new phase-difference frame is converted to a ΔT map using the PRFS equation, then added to body temperature (typically assumed ~37°C at baseline) to yield an absolute temperature value per voxel. This temperature map is color-coded — commonly blue for cool baseline tissue, cyan/green near the target temperature, and red for hot spots exceeding the plan — and overlaid semi-transparently on the anatomical magnitude image so clinicians can see temperature in direct spatial context with organ boundaries, vasculature, and the tumor outline.
Because thermal damage accumulates nonlinearly with both temperature and time, many systems also compute a running thermal dose map (CEM43, cumulative equivalent minutes at 43°C) alongside the instantaneous temperature map, integrating exposure over the whole session rather than showing only a momentary snapshot.
The temperature map's clinical value lies in what it reveals that could never be assumed from the applied power settings alone: tissue is heterogeneous, blood flow perfuses heat away unevenly (a phenomenon called the "heat sink" effect near large vessels), and small patient movements shift where energy actually couples into tissue relative to where it was planned.
A hot spot — a voxel cluster exceeding the safe upper temperature bound — risks thermal injury to healthy tissue or vasculature and must be identified within one or two update cycles. A cold region — undertreated tumor tissue lagging behind the 40–43°C target — risks a subtherapeutic dose that fails to sensitize the tumor to concurrent radiotherapy or chemotherapy, which is the primary reason hyperthermia is prescribed in the first place.
This live map is the direct input to the closed-loop control algorithm in the next stage — without it, the system would be running blind, relying only on a static pre-treatment plan with no way to detect or correct for how the actual patient anatomy responds to heating in real time.
This is the stage where MR thermometry stops being a passive monitoring tool and becomes an active steering signal. A control algorithm continuously compares the measured temperature map against the desired target pattern and computes the power and phase corrections that each applicator element needs to make, converting an imaging measurement directly into an actuator command.
The core of the closed-loop algorithm borrows directly from classical control theory. For each voxel or control point, the algorithm computes an error signal e(t) = T_target − T_measured(t), and a proportional-integral (PI) controller produces a correction:
u(t) = Kp · e(t) + Ki · ∫e(t)dt
where Kp (proportional gain) determines how aggressively the system reacts to the instantaneous deviation, and Ki (integral gain) accumulates persistent error over time to eliminate steady-state offset — for example, a region that consistently runs slightly cold due to a nearby blood vessel's cooling effect.
Because hyperthermia applicators have many independent elements (phased-array antennas or ultrasound transducer elements), this single-point control law must be extended to a multi-input multi-output (MIMO) problem: the algorithm maintains a sensitivity or "steering" matrix that maps each element's power and phase to its expected thermal contribution at each voxel, inverts or optimizes against that matrix, and distributes the correction across all elements simultaneously rather than adjusting one at a time.
Higher gain settings make the system correct deviations faster but risk overshoot and oscillation if pushed too aggressively relative to the thermal and imaging update lag; lower gain converges more slowly but more smoothly — the classic control-engineering tradeoff between speed and stability.
Closed-loop MR-controlled hyperthermia has been demonstrated in academic research systems since the 1990s and 2000s, and equivalent real-time thermometry feedback is now integrated into commercial MRgFUS platforms — the same PI-style control principle, applied to focused ultrasound rather than RF or microwave hyperthermia.
Translating a full 2D or 3D error map into a small set of applicator commands requires spatial aggregation: the algorithm typically weights errors within the defined target volume most heavily, applies smoothing to avoid reacting to single noisy voxels, and enforces hard safety constraints (maximum per-element power, maximum voxel temperature) before any correction is sent to hardware.
The computed adjustment — how much to raise or lower each element's power, and how to shift its phase to redirect the constructive-interference focus of a phased array — is sent to the hyperthermia or focused-ultrasound driver hardware, closing the loop: measure → compare → adjust → repeat, once every MR update cycle for the remainder of the treatment session.
With the control algorithm's corrections computed, the hyperthermia or ultrasound system executes them: individual applicator elements ramp power up or down and shift phase, continuously steering the actual heating pattern toward the planned target distribution for the remainder of the session. This is the payoff of the entire feedback chain — a treatment that adapts to the patient in real time rather than following a fixed, pre-planned protocol blindly.
Historically, hyperthermia treatments were delivered open-loop: a plan computed power settings in advance from a simulation of the patient's anatomy, and those settings ran largely unchanged for the session, with only intermittent invasive thermocouple readings (if any) to spot-check temperature at a handful of discrete points. This approach is blind to the exact heterogeneity of real tissue, blind to patient motion, and blind to the heat-sink effect of nearby blood vessels — all of which can cause substantial, uncorrected deviation between planned and delivered temperature.
Closed-loop MR-controlled treatment fundamentally changes this: every applicator adjustment is grounded in a fresh, whole-volume, non-invasive temperature measurement rather than a static prediction or a sparse probe reading. Clinically, this translates into tighter temperature control around the therapeutic 40–43°C window, meaningfully lower risk of thermal injury to healthy tissue or vasculature from undetected hot spots, and better assurance that the tumor volume actually receives the thermal dose the treatment plan intended — directly improving the reliability of hyperthermia as a radio- and chemo-sensitizer.
The practical difference is visible directly in convergence speed: with low control gain, a detected deviation may take many update cycles to correct, while with well-tuned higher gain the same deviation converges toward the target within just a few MR update cycles — without overshooting into an unsafe hot spot.
Adaptive power steering is not a one-time correction — it runs continuously, once per MR update cycle, for the full duration of the hyperthermia session (commonly 60 minutes or longer). As tissue properties shift during treatment (perfusion changes, tissue coagulation in ablative applications, patient micro-motion), the loop keeps re-measuring and re-correcting, so the delivered heating pattern tracks the target pattern far more faithfully over the whole session than any single upfront plan could guarantee.
This real-time adaptability is precisely what has driven the integration of MR thermometry feedback into modern MR-guided focused ultrasound (MRgFUS) platforms and MR-compatible hyperthermia applicators — turning MRI from a purely diagnostic tool into an active participant in steering the therapy itself.