Robotic-arm linac with real-time image-guided tumor tracking — frameless SRS/SBRT
Conventional radiosurgery relies on a rigid stereotactic head frame — screwed to the skull — to guarantee sub-millimeter geometric accuracy. CyberKnife eliminates the frame entirely by tracking the target directly with X-ray image guidance, using either implanted gold fiducial markers (soft-tissue targets) or the patient's own bony anatomy (cranial and spinal targets) as a real-time positional reference.
The Gamma Knife and conventional linac-based SRS traditionally depend on a rigid stereotactic frame fixed to the skull with pins, which physically constrains the head and provides a fixed coordinate system for targeting. This guarantees accuracy but is invasive, uncomfortable, restricted to single-day cranial treatments, and cannot be used on the body.
CyberKnife (Accuray) instead pairs two ceiling-mounted diagnostic X-ray sources with floor-mounted amorphous-silicon detectors, positioned orthogonally around the treatment couch. Every few seconds to minutes, this stereoscopic imaging pair acquires a pair of X-ray projections, digitally compares them against reference digitally-reconstructed radiographs (DRRs) generated from the planning CT, and computes the six-degree-of-freedom offset (3 translations + 3 rotations) between the patient's current position and the plan.
Because tracking happens continuously throughout treatment rather than once at setup, immobilization only needs to be "good enough" (a lightweight mask or vacuum cushion) — the imaging system, not the hardware, enforces final accuracy.
Removing the invasive head frame allows CyberKnife treatments to be fractionated over multiple days — impossible with a frame that must be reapplied and re-registered on each visit — enabling hypofractionated schedules that spare normal tissue via radiobiological fractionation.
CyberKnife selects a tracking algorithm matched to the treatment site:
• Skull tracking (6D Skull): for cranial lesions, the system registers the bony skull contour directly from the live X-rays against the planning CT DRRs — no fiducials needed, since the skull is essentially rigid relative to intracranial targets.
• Xsight Spine tracking: for spinal lesions, vertebral body bone structure itself provides the tracking reference, again fiducial-free, exploiting the fact that spinal tumors move rigidly with the adjacent vertebrae.
• Fiducial tracking: for soft-tissue targets that move independently of bone — lung, liver, pancreas, prostate — 3–5 small gold or platinum fiducial markers (roughly the size of a grain of rice) are implanted percutaneously or bronchoscopically near the tumor days before planning. Because soft tissue lacks a rigid bony reference, these radio-opaque seeds give the X-ray imaging system a sharp, trackable surrogate for the tumor itself.
Each imaging cycle acquires two near-simultaneous 2D X-ray projections from orthogonal angles. Pattern-matching software locates the fiducials (or bony landmarks) in both images and triangulates their 3D position. Comparing this against the planned position yields a rigid-body transform that the robot applies to the next beam's aim point before it fires.
This image-to-plan registration is what ultimately replaces the frame's mechanical guarantee with a software-and-imaging guarantee — end-to-end system accuracy is validated to be well within 1 mm in phantom studies, the benchmark that makes frameless intracranial and extracranial radiosurgery clinically viable.
X-ray imaging cannot run continuously — every image adds radiation dose, and constant fluoroscopy would be clinically unacceptable. Accuray's Synchrony Respiratory Tracking System solves this by building a real-time mathematical model that correlates easily-measured external chest-wall motion with the internal tumor position sampled only intermittently by X-ray, letting the system predict tumor location every treatment moment even between imaging snapshots.
Lung and liver tumors move substantially with every breath — commonly 5–20 mm, occasionally more, predominantly along the superior-inferior axis but with a smaller anterior-posterior and lateral component too. A static treatment plan aimed at a single "average" position would systematically underdose the tumor at the extremes of its excursion while irradiating healthy tissue the tumor only occupies part of the time.
Early motion-management approaches used breath-hold or gating (only treating during a narrow respiratory window), but these prolong treatment and are hard for sicker patients to sustain reproducibly. Synchrony instead tracks the tumor continuously and moves the beam with it.
Lightweight reflective optical markers (LEDs or passive spheres) are placed on the chest or abdominal surface and tracked at high frame rate (tens of Hz) by an overhead camera array — this is cheap, dose-free, and continuous. Separately, the X-ray imaging system periodically captures the true internal fiducial/tumor position.
Synchrony fits a mathematical correlation model — typically a polynomial relating external marker position (and its recent history/velocity) to internal target position, refit continuously as new X-ray samples arrive. Once trained, the model can predict internal tumor position at every instant from the external signal alone, dramatically reducing how often X-ray imaging is needed while still keeping tracking current.
The robot then continuously updates the linac head's aim along the predicted trajectory, and each new X-ray sample both corrects any drift and re-trains the model, keeping the prediction anchored to reality throughout the session.
The correlation model is what allows CyberKnife to treat moving extracranial tumors without breath-hold, abdominal compression, or gating — the beam follows the tumor rather than the tumor being forced to hold still for the beam.
More frequent X-ray sampling keeps the correlation model tightly anchored to the true internal position and shrinks real-time tracking error — but every additional image adds imaging dose on top of the therapeutic dose, and breathing patterns can drift (baseline shift, irregular depth) between samples. Clinical protocols balance imaging frequency against cumulative kV imaging dose, typically sampling every 15–60 seconds once the model is well-trained, with more frequent sampling only during initial model-building or when correlation confidence drops.
The defining hardware innovation of CyberKnife is its treatment head: a compact 6 MV linear accelerator mounted on the wrist of an industrial 6-axis robotic manipulator, the same class of arm used in automotive manufacturing, repurposed and validated for sub-millimeter medical positioning. This lets CyberKnife aim beams from hundreds of distinct positions in a roughly spherical volume around the patient, rather than being confined to a single mechanical isocenter.
A conventional gantry-based linac rotates around a single fixed mechanical isocenter — every beam in the plan passes through that one point in space, and conformality is achieved by shaping each beam's aperture (MLC) and arc trajectory around that fixed pivot.
CyberKnife abandons the fixed isocenter altogether. Its treatment-planning system selects beams from a pre-computed set of ~100–200 candidate "nodes" — robot positions and orientations distributed over a roughly spherical shell around the patient — and optimizes which subset of nodes, with which beam weights and collimator apertures, best conforms the dose to the target shape while sparing nearby organs at risk. Beams need not all intersect at one point; instead they converge geometrically on the target volume from many independent directions, allowing the dose cloud itself to take on essentially arbitrary, non-spherical shapes.
Because beams are optimized individually rather than constrained to a shared isocenter, non-isocentric planning can wrap dose tightly around irregular or concave target shapes — such as a tumor abutting the spinal cord — with a conformality and dose-falloff steepness that is difficult to match with a single rotating gantry.
Between beams, the robotic arm physically repositions the compact linac head to the next selected node — typically taking on the order of a second per move — then the imaging/tracking system re-verifies the target position before the beam fires (see Stage 4). This sequential, one-beam-at-a-time delivery through many nodes is what gives CyberKnife its geometric flexibility, but it is also the source of its principal trade-off versus Gamma Knife or gantry-based linac SRS: because each of the ~100–200 beams is delivered individually with imaging verification between many of them, total treatment time per fraction is typically longer (often 30–90 minutes) than single-isocenter systems that deliver dose through continuous arcs.
At each node the beam can be shaped by fixed circular collimators (5–60 mm diameter) or, on newer systems, the Iris Variable Aperture Collimator — a set of adjustable tungsten segments that changes aperture size electronically between beams without a physical collimator swap, or a multileaf collimator (InCise MLC) for more complex conformal apertures. Combined with hundreds of achievable beam directions, this beam-shaping flexibility is what allows the treatment planning system to sculpt dose distributions that tightly match irregular tumor geometry while achieving steep dose gradients toward adjacent critical structures.
Positioning the robot precisely at a node is only half the problem — the target itself is still moving. CyberKnife closes the loop by re-imaging the patient throughout delivery and adjusting the aim of each upcoming beam to the target's verified current position, rather than firing blindly at wherever the target was when the plan was made.
Without real-time correction, a beam aimed using only the last available image would systematically miss a target that has since moved — most visibly with breathing tumors, but even skull and spine targets can drift slightly from swallowing, small patient motion, or couch settling over a long fraction. CyberKnife's tracking loop continuously compares the Synchrony-predicted (or directly imaged) current target position against the position assumed at planning, and feeds that offset to the robot before every beam. The robotic arm has sufficient positioning freedom to apply small real-time corrective adjustments to its aim without needing to physically re-plan or re-select nodes.
Because each X-ray pair adds imaging dose, the system does not re-image before every single beam; instead it interleaves periodic verification images with model-predicted tracking (Stage 2) between them, re-imaging more frequently if the correlation model's confidence drops or the patient's breathing pattern changes. If imaging frequency is set too low relative to how quickly the target drifts, the gap between verified positions widens and the effective tracking error grows — beams increasingly rely on extrapolation rather than measurement. Conversely, very frequent imaging tightens tracking error but adds cumulative kV dose, so clinical protocols tune imaging cadence to the specific patient's motion amplitude and regularity.
Published end-to-end accuracy studies (imaging + correlation + robotic correction combined) report total system tracking errors under roughly 1.5 mm RMS for moving targets — a level of precision that approaches the intrinsic accuracy of frame-based systems despite tracking a target that never stops moving.
In practice, tracking correction is not a single big adjustment but a continuous small-scale steering process: the predicted target trajectory from the correlation model nudges the robot's aim frame-by-frame between images, and each new X-ray sample provides a correction snap back onto the verified true position, preventing the small prediction errors that accumulate between images from growing unchecked. The net effect, visualized against a hypothetical "no tracking" beam that simply fires at a stale, previously-imaged position, is a small but clinically important geometric miss that real-time tracking is specifically designed to eliminate.
After all beam nodes have fired — each individually verified and re-aimed onto the tracked target — their combined dose forms a single composite distribution. Because every beam was aimed at the target's true, moving position rather than a fixed point, the resulting dose cloud conforms tightly to the tumor's actual tracked volume, with a fall-off gradient steep enough to spare immediately adjacent normal tissue, all without any rigid immobilization frame.
The final dose distribution is the superposition of every individual beam fired during the fraction, each contributing a thin path of dose converging on the target from a distinct robotic node. Because tracking corrected each beam's aim to the tumor's true position at the moment of firing — rather than to a single planned isocenter — the composite dose envelope hugs the tumor's actual, breathing-inclusive tracked volume rather than a larger, motion-padded margin that would otherwise be needed to guarantee coverage of a target known to move.
This is the clinical payoff of everything upstream: frameless setup, respiratory correlation, non-isocentric robotic delivery, and continuous tracking correction all combine to let the treated volume shrink toward the true tumor size, sparing more of the surrounding healthy lung, liver, or other normal tissue than a static, larger-margin plan would allow.
CyberKnife's sub-millimeter, image-verified accuracy is what enables stereotactic hypofractionation — delivering the full ablative dose in just 1 to 5 large fractions instead of the 20–40 small fractions typical of conventional radiotherapy. This is only radiobiologically and clinically safe because the treatment team can trust that each large fraction actually lands on the tumor and not on adjacent normal tissue; the entire tracking architecture described in Stages 1–4 exists to make that trust justified for cranial SRS (single fraction) and extracranial SBRT (typically 3–5 fractions for lung, liver, pancreas, and spine).
CyberKnife's frameless, continuously-tracked approach trades one set of strengths for another compared to alternative radiosurgery platforms:
• Gamma Knife: uses ~192 fixed Cobalt-60 sources converging on a single isocenter with a rigid head frame (or frameless mask systems on newer units) — extremely fast dose delivery and very high intracranial precision, but limited to cranial targets only and traditionally frame-dependent for single-day treatment.
• Gantry-based linac SRS (e.g., fixed-isocenter VMAT/arc SRS): delivers dose through continuous rotating arcs around one isocenter — much faster per fraction than CyberKnife's sequential node-by-node delivery, and can treat extracranial sites with appropriate immobilization and gating, but generally offers less non-isocentric shaping flexibility and typically relies on less frequent intra-fraction imaging than CyberKnife's continuous tracking.
• CyberKnife: uniquely combines a frameless workflow, extracranial capability for moving soft-tissue targets, and continuous real-time tracking correction — at the cost of longer per-fraction treatment times (30–90 minutes) because dose is built up sequentially, one verified beam at a time, through many robotic positions rather than delivered through fast continuous arcs.
The choice between platforms is rarely about which is universally "better" — it is about matching a delivery philosophy (fast fixed-isocenter arcs vs. slower but highly flexible non-isocentric tracked delivery) to the target's location, size, shape, and whether it moves with respiration.