Truly painless cancer treatment, realized through medical image analysis.
Clinical trials and peer-reviewed papers demonstrate its accuracy, supporting reliable position verification and tumor tracking.
Clinical Evidence
PIL’s markerless tumor tracking is backed by data measured in clinical trials and clinical operation at QST Hospital (formerly the National Institute of Radiological Sciences). The key figures and peer-reviewed papers are summarized here for R&D leaders to use in internal review.
0.5mm
Tumor position accuracy with internal (markerless) gating (95% CI)
4.1 mm with surface-based gating
97.3%
Share of beam-on time with position error below 1 mm
200+
Cases treated at QST Hospital
(IGRT including markerless tracking)
<30ms
Tracking computation per frame
(lung ~29 ms, liver ~23 ms)
How markerless tracking works
Figure 1. The tumor appearance is learned from the planning 4DCT; during delivery the tumor position is estimated every frame from two-view fluoroscopy and the beam is on only inside the planned window. No fiducial markers are implanted.
Why surface-based gating is not enough: studies of external surrogates vs tumor position
Figure 2. Body-surface motion (external sensor) and tumor position differ in amplitude and phase even within the same breath, and the relationship changes from patient to patient and day to day. Gating on the surface signal can therefore deliver beam while the tumor is outside the planned window (conceptual diagram).
LUNG
Correlation between tumor motion and external surrogates (abdominal displacement, respiratory volume) ranged from r = 0.39 to 0.99 across patients, with phase offsets of −0.65 to +0.50 s. Of five patients measured on multiple days, only one showed a consistent relationship. The authors concluded that surrogate-guided treatment “might result in geographic miss”.
Int J Radiat Oncol Biol Phys 2004;60(4):1298-1306. doi:10.1016/j.ijrobp.2004.07.681
LUNG / LIVER / ESOPHAGUS
In 26 patients, the respiratory-sensor waveform was compared with 3D tumor motion. A phase shift of 0–0.3 s between the external waveform and tumor motion was observed, and additional margins were recommended for gating based on external sensors.
Int J Radiat Oncol Biol Phys 2004;60(3):951-958. doi:10.1016/j.ijrobp.2004.06.026
LIVER
Internal fiducials and external markers were measured simultaneously in four liver patients. Although correlated overall, tumor position varied by 2–9 mm for the same external-marker position, and the tumor-to-marker motion ratio ranged from 0.85 to 7.1. Margins should not rely on external markers alone.
Int J Radiat Oncol Biol Phys 2005;61(5):1551-1558. doi:10.1016/j.ijrobp.2004.12.013
CARBON-ION / DIRECT COMPARISON
Using treatment data from the same 10 patients, gating accuracy of markerless tracking of the tumor itself (internal gating) was compared with surface-based gating applied to the same data: 0.5 mm with internal gating versus 4.1 mm with surface-based gating (95% CI).
Int J Radiat Oncol Biol Phys 2016;95(1):258-266. doi:10.1016/j.ijrobp.2016.01.014
Peer-reviewed publications01–05: research at QST (NIRS) with industry partners that forms the origin of PIL’s technology. 06: reference (international benchmark)
01 — CLINICAL TRIAL / CARBON-ION
Carbon-Ion Pencil Beam Scanning Treatment With Gated Markerless Tumor Tracking: An Analysis of Positional Accuracy
Int J Radiat Oncol Biol Phys 2016;95(1):258-266. doi:10.1016/j.ijrobp.2016.01.014
The first clinical trial combining markerless tumor tracking with amplitude-based respiratory gating in scanned carbon-ion therapy, in 10 patients with lung and liver tumors. Tumor position was detected in real time with paired X-ray fluoroscopy and the beam was delivered only while the CTV was inside the PTV. Positional accuracy, imaging dose and throughput were evaluated.
Key results
- Internal gating accuracy 0.5 mm (95% CI)4.1 mm if external gating had been applied
- TRE < 1 mm in 97.3% of treatments
- Patient setup error 1.1 ± 1.2 mm / 0.6 ± 0.4°
- Fluoroscopic dose 23.7 ± 21.8 mGy per beambelow 487.5 mGy per treatment course
02 — MACHINE LEARNING / LUNG
A machine learning-based real-time tumor tracking system for fluoroscopic gating of lung radiotherapy
Phys Med Biol 2020;65:085014. doi:10.1088/1361-6560/ab79c5
A machine-learning method (extremely randomized trees) trained on DRRs generated from planning 4DCT, producing a tumor likelihood map on fluoroscopic images to estimate tumor position without fiducial markers. Evaluated on fluoroscopic data (15 fps) from eight lung cancer patients for accuracy and computation time.
Key results
- Tracking accuracy 1.03 ± 0.34 mm (mean ± SD)95th percentile 1.76 ± 0.71 mm
- Computation time 28.66 ± 1.89 ms per framereal-time tracking achieved
03 — REGRESSION MODEL / LIVER
Regression model-based real-time markerless tumor tracking with fluoroscopic images for hepatocellular carcinoma
Physica Medica 2020;70:196-205. doi:10.1016/j.ejmp.2020.02.001
For hepatocellular carcinoma, where the tumor is hard to see on fluoroscopy, the positional relationship between the diaphragm and the tumor is learned from 4DCT as a regression model, and tumor position is estimated from the diaphragm on fluoroscopic images. Evaluated on seven liver cases and 15 fluoroscopic sequences for diaphragm detection error, modeling error, tracking error and computation time.
Key results
- Tumor tracking error 1.30 ± 0.54 mmdiaphragm detection error 0.57 ± 0.62 mm
- Computation time 23.2 ± 1.3 ms per frame (tracking)training 69.0 ± 4.6 ms per frame
04 — COMMISSIONING / ROTATING GANTRY
Commissioning of a fluoroscopic-based real-time markerless tumor tracking system in a superconducting rotating gantry for carbon-ion pencil beam scanning treatment
Med Phys 2019;46(4):1561-1574. doi:10.1002/mp.13403
Final quality assurance before clinical use of the markerless tumor tracking system mounted on a superconducting rotating gantry: tracking accuracy with a moving chest phantom (machine learning and multi-template matching), four interlock functions, and gating / beam latencies.
Key results
- Tracking accuracy <0.52 mm (95% CI)all gantry angles and frame rates (MTM)
- Gate on/off latency <82.7 ± 7.6 ms
- Gating control system latency <3.1 ± 1.0 ms
- Beam irradiation latency <8.7 ± 1.2 ms
- Interlock signal delaystumor velocity 44.7 ms (1.3 frames) / image brightness 34 ms (1.0 frame) / two-view inconsistency within 5.0 ms
- Anomaly-detection interlock reduced tracking error 2.27 → 0.25 mm
05 — PATIENT-SPECIFIC DEEP LEARNING / LUNG
Real-time markerless tumour tracking with patient-specific deep learning using a personalised data generation strategy: proof of concept by phantom study
Br J Radiol 2020;93:20190420. doi:10.1259/bjr.20190420
A patient-specific deep-learning approach that trains a convolutional neural network for each lesion from many DRRs generated from the patient’s own planning 4DCT, avoiding the need for a large patient dataset. The gap between training DRRs and real X-ray images is bridged with random contrast transformation and noise addition. Validated with a digital phantom and an epoxy phantom.
Key results
- Tracking accuracy (frames with error <1 mm): 100% in simulation
- Phantom study: 100% for a 3 cm sphere, 94.7% for a 2 cm sphere
- Processing 32.5 ms per frame (30.8 fps)real-time tracking at 30 fps
06 — REFERENCE / INTERNATIONAL BENCHMARK
The markerless lung target tracking AAPM Grand Challenge (MATCH) results
Med Phys 2022;49(2):1161-1180. doi:10.1002/mp.15418
An international benchmark of markerless lung target tracking sponsored by the American Association of Physicists in Medicine. Using a common thorax phantom and patient-measured breathing traces, commercial systems (CyberKnife, Radixact, Vero, C-RAD) and preclinical approaches were compared with the same metric (percentage of 3D tracking error within 2 mm). PIL’s CTO took part as a co-author of the challenge.
Key results
- 30 institutions registered, 15 submissions4 in silico, 11 experimental
- 3D tracking error within 2 mm: 50–92% (in silico), 39–96% (experimental)
- Several approaches reached sub-millimeter accuracya common methodology for clinical adoption of markerless tracking
* Figures are as reported in each paper. The number of clinical cases (200+) is PIL’s own count. Full texts are available from the publishers. Contact us for an extended publication list or clinical trial details.
← Back to Why PIL?Contact
Let's solve medical imaging,
together.
Joint research, product development support, careers — feel free to get in touch.