The Radiation Oncology Feed For radiation oncologists · physicists · RTTs · dosimetrists
technical-relevance mrgrt motion-management

Training-free MRI tracking achieved submillimetre motion accuracy in liver metastases

ETLD tracked 106,000 cine-MRI frames with submillimetre error, while integrated segmentation achieved a mean global Dice score of 87.7%.

RadOncToday · ·1 min read

KEY POINTS

  • The study evaluated a training-free enhanced Tracking-Learning-Detection algorithm combined with an improved Chan-Vese segmentation model in 10 patients with colorectal liver metastases treated on an Elekta Unity MR-linac.
  • The original dataset contained 112,800 cine-MRI frames from 77 treatment fractions. After image-quality filtering and expert verification, 106,000 frames across 2,120 sequences remained for analysis.
  • Mean target-motion amplitude was 5.3 ± 4.5 mm. Across all patients, ETLD achieved mean absolute tracking errors below 0.8 mm, root-mean-square errors below 1.1 mm, correlation coefficients above 94.3%, precision above 99%, and recall above 98%.
  • The enhanced method had no complete tracking failures, compared with 158 failed sequences using the original TLD implementation. Approximately 75.1% of individual frames had displacement errors below the native MRI pixel spacing.
  • Integrated ETLD+ICV segmentation produced global Dice scores above 82% in every patient, with a mean of 87.7%. Mean frame-level Dice was 83.5 ± 17.4%, versus only 41.1% when the first-frame segmentation was simply carried forward without motion tracking.
  • Tracking ran at approximately 22 ± 7 frames/s, segmentation at 25 ± 16 frames/s, and total processing latency remained below approximately 100 ms per frame, without GPU acceleration.
  • On the external TrackRad benchmark, ETLD+ICV achieved a mean Dice score of 73.5 ± 5.3%, compared with 58.4 ± 5.8% for TransMorph and 86.3 ± 4.5% for SAM2. ETLD+ICV was fastest at 17 ± 6 frames/s, versus 11 ± 1 for TransMorph and 2 ± 0 for SAM2.
  • A critical limitation is that the investigators tracked blood vessels as surrogates for tumour motion because tumour boundaries were difficult to delineate reliably on the cine-MRI sequences. The study therefore does not yet establish tumour-tracking accuracy, gating performance, or dosimetric benefit during actual beam delivery.

CLINICAL TAKEAWAY

A lightweight, training-free algorithm can achieve genuinely real-time submillimetre MRI motion tracking without specialized GPU hardware. The external validation is encouraging, but direct tumour tracking and prospective integration with beam gating or tracking are required before clinical implementation.

SOURCE

Radiation Oncology