Blöcker Tom, Lombardo Elia, Marschner Sebastian N, Belka Claus, Corradini Stefanie, Palacios Miguel A, Riboldi Marco, Kurz Christopher, Landry Guillaume
Department of Radiation Oncology, LMU University Hospital, LMU Munich, Munich, Germany.
German Cancer Consortium (DKTK), Partner Site Munich, a Partnership between DKFZ and LMU University Hospital, Munich, Germany.
Phys Med Biol. 2024 Dec 24;70(1). doi: 10.1088/1361-6560/ad9dad.
. 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.. The first approach used a point-tracking model that propagates points from a reference contour. The second approach used a video-object-segmentation model, based on segment anything model 2 (SAM2). Both approaches were evaluated and compared against each other, inter-observer variability, and a transformer-based image registration model, TransMorph, with and without patient-specific (PS) fine-tuning. The evaluation was carried out on 2D cine MRI datasets from two institutions, containing scans from 33 patients with 8060 labeled frames, with annotations from 2 to 5 observers per frame, totaling 29179 ground truth segmentations. The segmentations produced were assessed using the Dice similarity coefficient (DSC), 50% and 95% Hausdorff distances (HD50 / HD95), and the Euclidean center distance (ECD).. The results showed that the contour tracking (median DSC0.92±0.04and ECD1.9±1.0 mm) and SAM2-based (median DSC0.93±0.03and ECD1.6±1.1 mm) approaches produced target segmentations comparable or superior to TransMorph w/o PS fine-tuning (median DSC0.91±0.07and ECD2.6±1.4 mm) and slightly inferior to TransMorph w/ PS fine-tuning (median DSC0.94±0.03and ECD1.4±0.8 mm). Between the two novel approaches, the one based on SAM2 performed marginally better at a higher computational cost (inference times 92 ms for contour tracking and 109 ms for SAM2). Both approaches and TransMorph w/ PS fine-tuning exceeded inter-observer variability (median DSC0.90±0.06and ECD1.7±0.7 mm).. This study demonstrates the potential of foundation models to achieve high-quality real-time target tracking in MRgRT, offering performance that matches state-of-the-art methods without requiring PS fine-tuning.
本研究旨在评估基于基础人工智能模型的两种用于磁共振成像（MRI）引导放疗（MRgRT）的实时目标跟踪方法。第一种方法使用了一种点跟踪模型，该模型从参考轮廓传播点。第二种方法使用了基于段分割模型2（SAM2）的视频对象分割模型。对这两种方法进行了评估，并相互比较，同时还与基于观察者间变异性以及基于变压器的图像配准模型TransMorph进行了比较，包括有无患者特异性（PS）微调的情况。评估是在来自两个机构的2D电影MRI数据集上进行的，这些数据集包含33名患者的扫描，有8060个标记帧，每个帧有2至5名观察者的注释，总共29179个地面真值分割。使用Dice相似系数（DSC）、50%和95%豪斯多夫距离（HD50/HD95）以及欧几里得中心距离（ECD）对生成的分割进行评估。结果表明，轮廓跟踪（中位数DSC 0.92±0.04和ECD 1.9±1.0毫米）和基于SAM2的方法（中位数DSC 0.93±0.03和ECD 1.6±1.1毫米）产生的目标分割与未进行PS微调的TransMorph相当或更优（中位数DSC 0.91±0.07和ECD 2.6±1.4毫米），略逊于进行了PS微调的TransMorph（中位数DSC 0.94±0.03和ECD 1.4±0.8毫米）。在这两种新方法之间，基于SAM2的方法在计算成本较高的情况下表现略好（轮廓跟踪推理时间为92毫秒，SAM2为109毫秒）。两种方法以及进行了PS微调的TransMorph都超过了观察者间变异性（中位数DSC 0.90±0.06和ECD 1.7±0.7毫米）。本研究证明了基础模型在MRgRT中实现高质量实时目标跟踪的潜力，提供了与最先进方法相匹配的性能，且无需PS微调。
