Label-Scarce Ultrasound Biopsy Needle Identification Using Dynamics-Guided Prior Modeling

Background/Objectives: Ultrasound-guided biopsy is essential for accurate needle placement in minimally invasive procedures. However, the limited availability of labeled medical data significantly constrains the development of needle segmentation models, and existing semi-supervised approaches rely solely on data-driven consistency constraints without incorporating domain-specific dynamic properties of the imaging process. We propose a semi-supervised unrolling optimization framework that combines dynamics-based data consistency with self-training. Methods: (1) A null-space projection-based Analytic Model (A) layer is unrolled into the network to suppress background interference and enforce dynamics-based constraints derived from temporal ultrasound sequences, in which the background subspace is estimated via singular value decomposition (SVD) of consecutive frames and the needle signal is recovered by projecting the observed image onto the orthogonal complement of this subspace. (2) A self-training strategy leverages unlabeled ultrasound data by generating soft pseudo-labels from a pretrained teacher network to iteratively supervise the student model, which shares the same unrolling architecture and is jointly trained with a supervised loss on labeled data and a pseudo loss on unlabeled data. Results: Evaluation on an ultrasound needle segmentation dataset demonstrated that the proposed method achieved a Dice score of 0.7429±0.0489 and an IoU of 0.5967±0.0917, outperforming the supervised UnrollNet baseline (Dice: 0.7298±0.0518), Co-Training (U-Net) (Dice: 0.6104±0.0949), and Meta Pseudo Labels (Dice: 0.6426±0.1140). The proposed method also yielded a mean trajectory angle error of 0.66∘±0.63∘ and a tip position error of 0.95mm±0.74mm. Conclusions: These results indicate the potential of the proposed framework for ultrasound-guided needle tracking.