ESTRO 2024 - Abstract Book

S4286

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2024

1690

Proffered Paper

Comparing Transformer training strategies for real-time tumor tracking in MRI-guided radiotherapy

Elia Lombardo 1 , Laura Velezmoro 1 , Sebastian Marschner 1,2 , Michael Reiner 1 , Stefanie Corradini 1 , Claus Belka 1,2,3 , Marco Riboldi 4 , Christopher Kurz 1 , Guillaume Landry 1 1 LMU University Hospital, Department of Radiation Oncology, Munich, Germany. 2 German cancer consortium (DKTK), a partnership between DKFZ and LMU University Hospital Munich, partner site Munich, Germany. 3 Bavarian cancer research center, (BZKF), Munich, Germany. 4 Ludwig-Maximilians-Universität München, Department of Medical Physics, Faculty of Physics, Garching, Germany

Purpose/Objective:

MR-linacs currently allow for intra-fractional motion management by means of gated beam delivery, and some groups are focusing on developing MLC-tracking [1]. Regardless of the beam adaptation strategy, precise real-time localization of the tumor or a surrogate tracking target on the 2D cine MRI is critical. Accurate tracking increases duty cycle efficiency during gating in the presence of breath-holds or allows for deformable MLC-tracking in free-breathing. Transformer models are by now considered the state-of-the-art deep learning algorithms in many applications. In this study, we leveraged TransMorph [2], a transformer which had been successfully applied for image registration of static 3D MRIs, for real-time deformable image registration of 2D cine MRIs. We trained the model with different unsupervised (Unsup), supervised (Sup) and patient-specific (PS) strategies and evaluated it by comparing warped contours against manually generated ground truth.

Material/Methods:

About 1 400 000 unlabeled raw cine MRI frames of 219 patients were collected at a single institution (abdomen, liver, lung, mediastinum, pancreas, prostate). These images were used for unsupervised training of TransMorph (TM Unsup). No image augmentations and a combination of the mean squared error (MSE) and the diffusion deformation vector field (DVF) regularizer as a loss function were used to train the model. Additionally, 7500 unlabeled raw cine MRIs of 35 patients were collected (same treatment sites as for the other set plus two heart tumor patients) from the same institution. For these patients, the tumor or a surrogate tracking target was manually segmented for all frames by two different observers to generate the ground truth labels for model optimization and evaluation. These patients were then partitioned into labeled fine-tuning (12), validation (10) and testing (13) sets. The labeled fine-tuning and validation set were used to optimize the pre-trained TM-Unsup in a supervised fashion, using image augmentations and a combination of the Dice Similarity Coefficient (DSC), MSE and diffusion regularizer as loss function (TM-Sup). Finally, we again started from the pre-trained TM-Unsup to train patient-specific models based on randomly selected 8 frames from the beginning of the cine MRI of each patient (TM-PS). For this, we used the same image augmentation and loss function as for the supervised model. For all three models, we selected the first frame of a cine and its corresponding contour as moving images and the rest of frames as fixed images for the registration. All cine MRIs were acquired during treatments in the sagittal plane with either 4 Hz or 8 Hz imaging frequency and in-plane spatial resolutions of 2.4×2.4 or 3.5×3.5 mm 2 . We resampled all frames to 1×1 mm 2 and cropped to 224×224 pixels. The