Answers are grounded in this paper and cite their sources.
Inter-frame probe motion, cumulative pose error and long-range spatial drift are problems that challenge the accurate three-dimensional reconstruction of freehand two-dimensional ultrasound sequences. In this work, a probe-motion-sensitive spatial-temporal deep learning architecture was created, which incorporates motion-sensitive consecutive-frame representation, residual spatial feature extraction, bidirectional temporal modelling, temporal attention, and long-horizon residual refinement. The framework was tested on subject-independent TUS-REC2024 ultrasound scans in terms of trajectory- and landmark-based measures. Compared to the original model, the proposed motion-sensitive model decreased the absolute trajectory error RMSE by 95.74 mm to 52.94 mm and the mean target registration error by 85.86 mm to 44.43 mm. The long-horizon refinement further decreased the mean target registration error to 40.68 mm with statistically significant mean paired improvement of 3.75 mm. The correlation between scan progression and registration error dropped to 0.513, which means that there is less spatial error accumulation at long sequences. Linear and S-shaped trajectories showed the greatest performance improvements, whereas curved probe motion was more difficult. The results show that explicit inter-frame motion modelling with temporal context can significantly enhance probe-motion estimation and spatial consistency, but there is still a non-negligible residual reconstruction error.
No takes yet. Share an insight, caveat, or question.
R et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: