OPT-former enabled near-real-time automated eye tracking during ocular proton therapy
OPT-former matched a three-network segmentation pipeline while processing eye images 2.6 times faster and improved further with patient-specific fine-tuning.
Tag
technical-relevance
uTPS produced deliverable Halcyon VMAT plans comparable with Eclipse
uTPS generated Halcyon VMAT plans with generally comparable dosimetry to Eclipse and greater than 97% gamma passing rates.
Deep learning predicted distant metastasis from CT and dose in head and neck cancer
A model integrating planning CT, 3D dose and tumor contours achieved strong external prediction of distant metastasis but weaker local-recurrence performance.
Robust optimization minimized lung-dose trade-offs in carbon-ion planning for stage I NSCLC
Robust optimization improved target coverage with less lung-dose penalty than margin expansion, while density override was particularly useful for solid tumors.
Nearly one-third of abdominal radiotherapy patients could not maintain planned DIBH
Despite successful simulation, 31.7% required DIBH gate modification or conversion to free breathing during abdominal radiotherapy.
Multimodal AI predicted radiation proctitis across different CT devices with 0.865 AUC
Combined radiomics, foundation-model and clinical features predicted grade ≥2 radiation proctitis with 0.865 AUC on an independent imaging device.
PTV coverage prescriptions poorly represented delivered tumor dose in lung SBRT
D95% and D98% PTV prescriptions produced substantial interpatient variability, while GTV- and ITV-based median or mean doses were more accurate.
Deep learning generated CT-free PET with near-zero SUV bias in NSCLC
Emission-only PET synthesis produced 0.72% SUVmax bias and 99.2% contrast recovery in an independent NSCLC test cohort.
CBCT-derived synthetic CT predicted head and neck replanning with 92% sensitivity and specificity
DART predicted clinically triggered head and neck adaptation with 92% sensitivity and specificity using a unified 2-Gy dosimetric threshold.
Reusable 3D-printed devices shaped small proton fields for under $1 in filament
Hollow 3D-printed proton collimators and compensators were produced within 45 minutes and accurately shaped small experimental proton fields.
Brainlab reference beam model enabled one-day SRS commissioning with accurate Monte Carlo calculations
Brainlab Elements commissioning was completed in one day, while Monte Carlo generally agreed with end-to-end measurements within 3%.
OCTAVIUS 4D achieved 98.3% gamma passing for MR-LINAC SBRT QA
Across 40 Unity SBRT plans, OCTAVIUS 4D achieved 98.3% mean local gamma passing with sub-degree angular agreement.
Deep learning reduced combined truncation and metal artifacts in half-detector CBCT
HD-TMAR improved metal-region PSNR from 25.5 to 35.5 dB in simulated half-detector CBCT while preserving dental anatomy.
MR-only CyberKnife planning achieved near-equivalent dosimetry for brain metastases
Synthetic CT reproduced CyberKnife target dosimetry closely and enabled submillimetre translational image guidance, although rotational errors were more variable.
Task-specific deep learning reduced CT metal artifacts in cervical brachytherapy
Task-specific deep learning halved pelvic artifact burden, improved applicator reconstruction to 0.1 mm, and shortened organ contouring by up to 40%.