Physics in Medicine and Biology, Год журнала: 2024, Номер 70(1), С. 015004 - 015004

Опубликована: Дек. 11, 2024

Abstract Objective . This study aimed to evaluate two real-time target tracking approaches for magnetic resonance imaging (MRI) guided radiotherapy (MRgRT) based on foundation artificial intelligence models. Approach The first approach used a point-tracking model that propagates points from reference contour. second video-object-segmentation model, segment anything 2 (SAM2). Both were evaluated and compared against each other, inter-observer variability, transformer-based image registration TransMorph, with without patient-specific (PS) fine-tuning. evaluation was carried out 2D cine MRI datasets institutions, containing scans 33 patients 8060 labeled frames, annotations 5 observers per frame, totaling 29179 ground truth segmentations. segmentations produced assessed using the Dice similarity coefficient (DSC), 50% 95% Hausdorff distances (HD50 / HD95), Euclidean center distance (ECD). Main results showed contour (median DSC 0.92 ± 0.04 ECD 1.9 1.0 mm) SAM2-based 0.93 0.03 1.6 1.1 comparable or superior TransMorph w/o PS fine-tuning 0.91 0.07 2.6 1.4 slightly inferior w/ 0.94 0.8 mm). Between novel approaches, one SAM2 performed marginally better at higher computational cost (inference times 92 ms 109 SAM2). exceeded variability 0.90 0.06 1.7 0.7 Significance demonstrates potential of models achieve high-quality in MRgRT, offering performance matches state-of-the-art methods requiring

Язык: Английский