سامانه همایش‌ها TSTA | ثبت‌نام و ارسال مقاله کنفرانس‌ها و کنگره‌های ملی

صفحه اصلی / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

CRAFT-Flow: Cross-Attentional Refinement for Robust Optical Flow Estimation in Cardiac MRI via Deep Learning

نویسندگان :

Hamed Aghapanah Roudsari¹ Reza Ashiri Gudarzi² Morteza Choubin³

1- دانشگاه علوم پزشکی اصفهان 2- دانشگاه ملایر 3- دانشگاه ملایر

کلمات کلیدی :

Cardiac MRI،Cross-Attention،Deep Learning،Motion Estimation،Optical Flow،Transformer Networks،Unsupervised Learning

چکیده :

Abstract— Optical flow estimation in medical imaging, particularly in dynamic cardiac MRI (CMRI), presents significant challenges due to complex non-rigid motion, low contrast, and absence of ground-truth labels. While deep learning has revolutionized optical flow in natural scenes, its application to medical sequences remains limited by model generalization and noise sensitivity. In this work, we propose CRAFT-Flow, a novel deep architecture that integrates cross-attentional transformers with a refined PWC-Net backbone to achieve robust, high-precision motion estimation in synthetic and real cardiac MRI sequences. Inspired by the CRAFT model’s success in large-displacement flow estimation, we redesign the correlation mechanism using cross-attention transformers, replacing classical correlation volumes to enhance long-range correspondence and reduce noise artifacts. Our model is trained on synthetic 4D cardiac phantoms generated via MRXCAT and XCAT, enabling supervision under realistic motion patterns. We further introduce a multi-scale warping and unsupervised refinement framework to adapt the model to unlabeled clinical CMRI data. Extensive experiments on Sintel, KITTI, and a custom cardiac dataset demonstrate that CRAFT-Flow achieves state-of-the-art performance, with a 21% reduction in EPE compared to prior methods and superior robustness under motion blur and noise. The proposed framework opens new pathways for motion-aware analysis in cardiac imaging, autonomous driving, and video understanding.