Can the scan used to aim radiotherapy become accurate enough for dose planning? A physics-aware model may do it without another CT scan

The scan used to aim radiotherapy helps position a patient, yet it often cannot be trusted for the dose calculation that

Radiotherapy needs more than a rough picture of where a tumor sits. It must know how much tissue lies between the skin and the target, because each layer absorbs or scatters energy in a different way. The treatment room scan is called cone-beam CT, or CBCT, and it helps staff line up the patient before every session. That job is useful, yet the same image often cannot be trusted when clinicians calculate how much dose will reach each organ.

Planning CT uses a rotating X-ray source and detector in a helical path, which gives cleaner numbers for tissue density. CBCT is built into the treatment room, so it can be taken quickly while the patient lies on the couch. The tradeoff is that its beam spreads through more air and body parts before reaching the detector. That extra scatter washes out fine differences between muscle, fat, and bone. Beam hardening happens when low-energy X-rays are absorbed first, leaving a harder beam that changes the measured darkness of tissue. Truncation occurs when part of the body falls outside the scanned field, so the image is rebuilt from incomplete information. Detector imperfections and reconstruction limits add streaks, blurring, or warped edges.

The numbers in a CT image are Hounsfield units, which describe how strongly tissue stops X-rays. Dose engines use those numbers to predict where radiation will deposit energy. When CBCT values drift, the predicted dose can be too high in one spot and too low in another. Soft tissue contrast falls, spatial distortions creep in, and small organs may appear shifted or blurred. Adaptive therapy wants to update plans as tumors shrink or organs move, but doing that from CBCT alone is risky. A cleaner CT equivalent image would let clinicians reuse the treatment room scan for planning without exposing the patient to another diagnostic scan.

How a rotation check borrowed from scanner physics steadies image rebuilding

When a scanner turns around a patient, each X ray beam makes a shadow of the body from one angle. If that same body is rotated by a small angle, the shadows do not become new shapes. They simply move along the detector by the matching amount. A CT machine knows this relationship because its source and detector follow a circular path, so the geometry gives a built in test for any volume it rebuilds.

The image rebuilding method uses that test as a training rule. First, a generative model tries to turn a noisy cone beam scan into a cleaner CT like volume. Then the rebuilt volume is rotated by small angles inside its plane. The software simulates what X ray shadows that rotated volume would cast, using the same kind of forward projection that describes how a scanner collects data. Those simulated shadows are compared with the real target CT projections after shifting them by the same angle. If the rebuilt anatomy cannot survive this rotation check, it means some tissue density or edge has been invented in a way that contradicts the scanner geometry.

This penalty does not ask the model to copy one picture more closely. It asks for consistency across angles, which is closer to how physical imaging works. The rule can discourage blobs, streaks, and soft tissue values that look plausible on screen but would produce impossible measurements when viewed from another direction. Because the check comes from the scanner’s own acquisition path, it gives the rebuilding process a stable anchor. The model still learns anatomy, contrast, and fine structure from paired scans, yet it is also taught to respect the geometry that created the data.

What improves when noisy scans are rebuilt across different machines and patients

Training on scans from two different hospitals makes the rebuilt images more reliable when they are tested on either hospital alone. On the phantom test set, peak signal to noise ratio rose from about thirty one decibels to thirty eight decibels. Structural similarity moved from zero point nine two to zero point nine nine. That means the rebuilt volume stayed much closer to the reference CT, both in brightness and in shape. The gain was largest when the model had seen more than one scanner setup. CycleGAN and a plain diffusion baseline improved too, but they did not reach the same fidelity. In mixed training, the rebuilt phantom images gained roughly seven decibels over the plain diffusion model. That difference is not a small tweak, because image comparison scores use a logarithmic scale. The clinical data showed the same direction, though with a smaller gain. There, peak signal to noise ratio improved by about one point eight decibels over the plain diffusion model. The improvement appeared across different regions of the head and neck slices. Hounsfield unit profiles also tracked the reference CT more closely, especially inside tissue boundaries. Visually, mixed training produced sharper anatomical edges and fewer intensity distortions. Dose calculation relies on stable tissue values, so these changes are useful. The model did not need a larger single dataset to get the benefit. It learned from diversity, including different patient groups and scanner related noise patterns. The result suggests that adding another acquisition environment helps the rebuilt image hold up under new conditions. Variability also fell, which points to steadier performance rather than a lucky best case. On the clinical test set, lower variability meant fewer slices drifted far from the reference. The rebuilt volumes therefore became more useful when moving between machines and patients.

Why the finished model can run quickly without extra projection steps

The finished model can turn a noisy cone-beam scan into a CT-like volume in about 58 seconds per patient, without asking the scanner for another set of measurements. That speed comes from two design choices. First, the network works in a compressed latent space rather than directly on every voxel of a full three-dimensional head scan. A small autoencoder first squeezes each volume into a compact representation that keeps depth information while reducing in-plane detail. The diffusion model then learns to clean and complete that smaller object, which is far cheaper than repeatedly editing millions of voxels. Second, the physics check that ties the output back to scanner geometry is applied during training, not at test time. While learning, it occasionally checks whether a candidate volume would produce scanner readings consistent with a cleaner CT scan. This happens only every tenth epoch, so its cost stays below 15 percent of total training time. Once training ends, that projection machinery is switched off. At inference, the model receives only the cone-beam input and runs a fixed sequence of 100 deterministic denoising steps in latent space. It does not solve a new scanner equation, search for missing rays, or wait for extra data from the machine. The result is a fast workflow that fits into a treatment room routine, where clinicians need an image soon after positioning.

A physics check that stays after training

The most useful idea here is to let the scanner geometry teach the model what a plausible image must do. During training, rotating a synthetic volume should move its projections in exactly the same way. If the generated image breaks that rule, the loss pushes it back toward physically consistent Hounsfield values. That constraint does more than sharpen edges. It helps the model keep tissue densities believable when it moves from one scanner setup to another. It also improves fidelity on both phantom data and clinical head and neck scans. The gains are concrete: a large jump in image similarity on the phantom set, a steadier improvement on patient data, and better Hounsfield accuracy inside tissue boundaries.

The broader promise is practical. If a treatment team can recover CT quality from the scan already used for positioning, adaptive radiotherapy may need fewer extra acquisitions, less radiation, shorter waits, and simpler planning. The method also runs quickly at inference because the physics operators are only needed while training, not during patient use. That makes it easier to imagine in a busy clinic than another slow reconstruction step.

The obstacles remain real. The evidence comes from paired head and neck data, so wider adoption will need larger cohorts, more body sites, and scanner types that differ even more strongly. Clinical trust also requires showing that the synthesized densities are safe for dose calculation across tumors, organs at risk, and changing anatomy over a treatment course. If an image can be rebuilt with enough confidence from the machine’s own measurements, what else in radiotherapy might stop being treated as disposable?