A Review of Deep Learning for Respiratory Motion Tracking in <scp>2D</scp> and <scp>3D</scp> Ultrasound Sequences for Image‐Guided Therapy

ABSTRACT Ultrasound (US) imaging is widely used in image‐guided therapies due to its real‐time capability, high temporal resolution, cost‐effectiveness, safety, and portability. While many US tracking methods exist, most reviews focus on conventional techniques and overlook recent advances in deep learning (DL). Existing DL surveys often emphasize modalities like MRI (Magnetic Resonance Imaging) or CT (Computed Tomography), but US tracking presents unique challenges, such as speckle noise, acoustic shadowing, and occlusions, that require tailored solutions. With increasing interest in automating US‐guided procedures, particularly for respiratory‐induced motion (RIM) tracking, this review fills a gap by focusing on recent DL‐based methods specifically developed for US motion tracking. We conducted a targeted review of state‐of‐the‐art DL techniques, categorizing them by network architecture. Special attention is given to studies using benchmark datasets like CLUST (Challenge on Liver Ultrasound Tracking), which enable standardized performance comparison. DL‐based methods demonstrate improved accuracy and robustness in difficult scenarios, such as out‐of‐plane motion and occlusions, where traditional approaches often underperform. DL has significantly improved the feasibility of markerless US tracking in clinical settings, supporting more automated and reliable image‐guided interventions. While conventional methods still offer value, hybrid strategies that integrate DL with traditional tracking or sensor fusion approaches show strong potential. This review complements prior surveys on conventional US tracking, serving as a useful resource for researchers developing robust, automated US‐based applications.