Real-Time Brachial Plexus Ultrasound Segmentation Using Lightweight Hierarchical Temporal Fusion

Identification of the brachial plexus on ultrasound remains difficult and operator-dependent. AI could assist with ultrasound-guided brachial plexus blocks, but real-time segmentation of nerve elements remains understudied. This study introduces a lightweight video segmentation framework for the sequential ultrasound imaging technique used in supraclavicular blocks. To reduce annotation workload, a semi-automated pipeline converts clinician-drawn bounding boxes into pixel-level masks using a tracking algorithm and UltraSam, an ultrasound-specific foundation model. The proposed model inserts a hierarchical temporal fusion module between the encoder and decoder of standard segmentation backbones. It employs a single layer of convolutional recurrent units with depthwise separable convolutions that connect hierarchical temporal features across multiple scales. We train with truncated backpropagation through time on long sequences and with domain-specific temporal-consistency and spatial-compactness losses. On a Sonosite test set of 24 videos, the SegFormer-B0-based model adds only 1.48 GFLOPs and 1.0 M parameters, yet improves the mean IoU from 40.53% to 43.32% and the $F_{1}$ score from 53.90% to 56.67% over the non-temporal baseline. Compared with SAM2 used as a tracker on dense whole-video ground truth, our autonomous model attains 4× higher mean IoU and a 3.6× lower temporal instability score. Cross-vendor validation on GE and Mindray devices shows that fine-tuning with 18 target videos yields consistent gains, and the model runs at 20.2/29.3 ms per frame on the Galaxy Tab S9/S8 mobile NPUs. The findings support safer ultrasound-guided regional anesthesia and could be adapted to other real-time medical image segmentation tasks.