Real-time patient dose verification during radiotherapy delivery — diodes, OSLDs, and MOSFETs on the skin catch delivery errors before they compound.
In-vivo dosimetry (IVD) begins before the beam ever turns on: small solid-state or luminescent detectors are taped directly onto the patient skin at the point where the beam central axis enters, and — for exit dosimetry — where it leaves. Each detector's response depends on its individual calibration factor, so correct placement and correct detector identity are as important as the measurement itself.
Three detector families dominate clinical in-vivo dosimetry, each with distinct readout physics:
• Silicon diodes: p-type or n-type semiconductor diodes generate electron-hole pairs proportional to absorbed dose; the resulting charge is read in real time through an electrometer during beam-on. Diodes are the fastest-responding option — usable for live, beam-on monitoring — but their sensitivity drifts with accumulated dose and temperature, requiring periodic recalibration (typically every 500–1000 cGy accumulated).
• Optically Stimulated Luminescent Dosimeters (OSLDs, Al2O3:C): trapped electrons released by laser stimulation emit light proportional to dose. OSLDs are read out after irradiation (not live), but they are small, largely energy-independent, reusable, and highly reproducible (±1.5% per reading), making them the preferred choice for point-dose verification and audits.
• MOSFET dosimeters: irradiation causes a permanent shift in threshold voltage of a metal-oxide-semiconductor transistor, read with a bias voltage after exposure. MOSFETs are extremely small (sub-mm sensitive volume), useful for surface and cavity measurements, but have a finite lifetime (limited total accumulated dose) before replacement.
Detector calibration factor N_D converts raw signal to dose: D = M × N_D × k_T,P × k_energy × k_SSD × k_fieldsize, where M is the raw reading and the k-factors correct for temperature/pressure, beam energy dependence, source-to-surface distance, and field size — each detector's correction factors are established individually against a calibrated ion chamber before clinical use.
Entrance dosimeters are placed on the skin along the central axis of the beam, at the point where it enters the patient, typically built up with a small cap of tissue-equivalent material to reach the depth of dose maximum (dmax) and eliminate contamination from low-energy scattered electrons. Exit dosimeters are placed on the opposite skin surface, directly in line with the beam, and measure the dose that has traversed the full patient thickness — a quantity influenced by patient thickness, tissue heterogeneities, and beam attenuation.
Together, entrance and exit readings allow reconstruction of the midplane (in-patient) dose using empirical or model-based algorithms, giving a full check of the dose actually delivered to the tumor depth, not merely the skin dose. Placement accuracy matters: a detector shifted a few millimeters off the central axis, or placed in a penumbra region, will read a dose that has nothing to do with the intended measurement point, and can generate a false discrepancy alarm.
As the linac delivers monitor units, each in-vivo detector converts incident radiation into an electrical or optical signal that is digitized, corrected, and converted into a dose value continuously or immediately after the beam turns off. For diode systems this happens live, allowing therapists to interrupt delivery if a gross error is detected mid-fraction.
During beam delivery, the detector's raw signal — charge collected for diodes and MOSFETs, or stored luminescence trap population for OSLDs read post-irradiation — accumulates as monitor units are delivered. An electrometer or dedicated in-vivo dosimetry system integrates this signal over the full field delivery, then applies the detector-specific calibration factor and correction factors to yield a dose in cGy.
D_measured(t) = ∫ S(t) dt × N_D × k_T,P × k_energy × k_wedge × k_SSD
For modulated deliveries (IMRT, VMAT), the signal is no longer a simple flat integration — dose rate, gantry angle, and MLC aperture all change continuously through the arc, so real-time systems must timestamp and log signal alongside machine parameters to reconstruct a fraction-level dose rather than a single static point value.
Several factors introduce noise or systematic offset into the acquired signal beyond the intended dose response:
• Angular dependence: diode response can vary by 1–3% with beam incidence angle, relevant for arc therapy and oblique fields. • Temperature drift: diode sensitivity changes roughly −0.3%/°C; skin-surface temperature differs from calibration-room temperature. • Dose-rate dependence: some MOSFETs show a small sensitivity change at very high instantaneous dose rates from flattening-filter-free beams. • Cable and connector artifacts: intermittent contact produces signal dropouts that look like large negative deviations if not flagged.
Modern in-vivo dosimetry systems apply these corrections automatically in software, but persistent outliers are flagged for the physicist to review before the measurement is accepted as valid.
Because diode response is read continuously during beam-on, gross delivery errors — e.g., a collimator jaw not opening, or an MU readout fault — can in principle be caught while the beam is still firing, giving therapists the chance to interrupt delivery before the full planned dose has been delivered incorrectly.
Once the fraction is delivered and the detector reading is converted to dose, the measured entrance and exit values are compared against the dose predicted by the treatment planning system (TPS) for that specific beam, field, and patient setup. This comparison is the core analytic step of in-vivo dosimetry — it converts a raw number into a pass/fail clinical decision.
The percentage deviation is calculated per fraction as:
Δ% = (D_measured − D_planned) / D_planned × 100
D_planned is extracted directly from the TPS dose calculation at the detector's equivalent depth and field geometry — not simply the prescription dose, since detector position, SSD, field size, and wedge/blocking all modify the expected reading at that specific point. A well-configured in-vivo dosimetry program pre-computes an expected reading for every field of every patient's plan before the first fraction, so the comparison at treatment time is immediate.
Small deviations (within ±2–3%) are expected from ordinary measurement uncertainty and normal day-to-day setup variation and require no action. Deviations that approach or exceed the institutional tolerance band trigger a closer look — first at the measurement itself, then, if the reading holds up, at the delivery.
Entrance dose is a relatively "clean" measurement: it depends mainly on field size, SSD, wedge/blocking factors, and detector calibration, all of which are well characterized. Exit dose additionally depends on total patient thickness, tissue heterogeneities (lung, bone, air cavities) along the beam path, and any daily variation in patient separation — so exit readings are inherently noisier and are typically given a wider tolerance band.
Because of this, many programs weight entrance dose more heavily as the primary machine/setup-error check, while treating a large exit-dose deviation with a normal entrance reading as a signal pointing toward anatomy change (e.g., weight loss, bladder/rectum filling, tumor shrinkage) rather than a machine or setup fault.
A field with a normal entrance reading but a exit dose 10% low is a classic signature of increased patient separation along the beam path — often from bladder or rectal filling in pelvic treatments, or unaccounted weight gain — rather than a delivery error at the machine.
When a measured dose falls outside the tolerance band, the discrepancy cannot simply be dismissed or corrected — it must be investigated and resolved before the next fraction, because an uncorrected systematic error compounds across the remainder of the treatment course. In-vivo dosimetry programs report catching clinically meaningful delivery errors in roughly 1–3% of all fractions measured.
A structured investigation follows a fixed order, cheapest and most likely causes first:
1. Measurement artifact: re-check detector placement, connector integrity, and whether the correct calibration factor was applied for that specific detector and field — a surprisingly common source of a "false" discrepancy.
2. Setup error: review the portal image or cone-beam CT against the reference — a couch shift, rotation, or immobilization device seated incorrectly changes the effective SSD and field alignment, directly changing entrance dose.
3. Patient anatomy change: compare today's CBCT separation to the planning CT — weight loss/gain, bladder or rectal filling, tumor regression, or edema all alter attenuation along the beam path and are a very common cause of exit-dose drift over a multi-week course.
4. Machine malfunction: check the linac's own QA logs (output constancy, MLC log files, wedge/blocking device confirmation) for that treatment session — output drift or a stuck jaw or wedge is rarer but more serious.
5. Wrong plan parameters: verify the delivered field matches the intended plan — field size, wedge angle, blocking, MU — a transcription or record-and-verify linkage error is uncommon but among the most serious causes because it may repeat every fraction until caught.
Because a wrong-plan or wrong-MU error repeats identically at every fraction until detected, in-vivo dosimetry's greatest clinical value is not any single flagged reading but the fact that it is checked at every fraction — the first fraction's IVD reading is the most important dosimetric check a patient receives, since it is the only chance to catch a systematic error before it compounds.
Once a cause is identified, the corrective action depends on the category: setup errors trigger re-training or additional imaging guidance for subsequent fractions; anatomy changes beyond a threshold trigger a replan; machine issues are escalated to physics/engineering for machine QA before further treatment; and plan-parameter errors require immediate correction of the record-and-verify system and, in the most severe cases, notification per the institution's medical event reporting policy (and potentially regulatory reporting, depending on the magnitude and jurisdiction).
Every investigated discrepancy is documented — cause, correction applied, and dose implication for the fraction — both to close the loop for that patient and to build the institutional dataset used to refine tolerance levels and detector calibration protocols over time.
A single fraction's deviation is a snapshot; the full value of in-vivo dosimetry emerges when every fraction's deviation is plotted across the entire course of 25–35 fractions. A trend line reveals patterns invisible in any one measurement — a slow systematic drift, a step change coinciding with a specific event, or simple random scatter around zero that requires no action at all.
Three distinct patterns appear in a fraction-by-fraction deviation plot:
• Random fluctuation: deviations scatter within roughly ±2–3% around zero from fraction to fraction with no directional trend — this is expected measurement and setup noise and requires no intervention, provided no individual point breaches the tolerance band.
• Systematic drift: deviations move gradually and consistently in one direction over many fractions — classically seen as a slow increase in exit-dose deficit over a course of pelvic or head-and-neck radiotherapy as the tumor shrinks and normal tissue changes shape, progressively altering attenuation. A run of 3 or more consecutive fractions trending the same direction is generally sufficient to flag for review even if no single fraction has yet crossed the tolerance line.
• Step change: a sudden jump in deviation that then remains roughly constant, most consistent with a discrete event — a new immobilization device, a change in patient positioning technique, or a corrected/re-corrected setup error — rather than a gradual physiological process.
When a systematic drift is confirmed and exceeds an institutional action threshold (commonly a sustained 3–5% shift), the response is typically an adaptive replan: a new CT is acquired, the treatment plan is recalculated on the patient's current anatomy, and subsequent fractions are delivered against the updated plan. This directly closes the loop that in-vivo dosimetry opened — the measurement that revealed the problem also verifies, over the following fractions, that the correction worked.
Aggregated across many patients, fraction-level trend data also feeds back into departmental quality assurance: tracking how often drift occurs by treatment site helps target which patient populations benefit most from mid-course re-imaging or replanning protocols, and confirms that detector calibration and correction factors remain valid across the full range of clinical field sizes, energies, and SSDs used in practice.
AAPM TG-62 and subsequent guidance frame in-vivo dosimetry not as a one-time check but as a continuous quality-assurance loop: every fraction is measured, every deviation is trended, and the trend — not any single reading — is what ultimately confirms that the dose the patient actually received matches the dose that was planned across the entire course.