EVENT DOSSIER
The researchers proposed a semi-supervised segmentation model combined with 6D planar pose regression for pipelines, aiming to provide continuous and real-time standard planar proximity feedback for fetal brain ultrasound scans. The model utilizes labeled SPs and unlabeled slices to process 3D ultrasound volume data, achieving an mIoU of 0.93 on SPs and 0.86 on non-SPs. A classification mechanism is integrated to filter out frames with no fetal brain images. The system is deployed on NVIDIA Clara AGX edge devices, achieving a real-time inference speed of 39 Hz, exceeding clinical collection standards. The study was retrospectively validated using real scan videos from 17 ultrasound physicians with different experience levels; it was found that operators tend to freeze near local minima of proximity rather than align precisely, and proximity itself cannot predict the quality score of expert SPs. This approach supplements existing technologies with sensor-based continuous distance detection, moving towards image navigation support in prenatal scans.
本文提出一种半监督分割模型与 6D 平面姿态回归相结合的管道，旨在为胎儿脑部超声扫描提供连续、实时的标准平面（SPs）邻近度反馈。该模型利用标注 SPs 及未标注切片处理 3D 超声体积数据，在 SPs 上达到 0.93 mIoU，在非 SPs 上达到 0.86 mIoU，并集成分类机制以过滤无胎儿脑图像帧。系统部署于 NVIDIA Clara AGX 边缘设备，实现 39 Hz 的实时推理速度，超过临床采集标准。研究基于 17 名不同经验水平的超声医师的真实扫描视频进行回顾性验证，发现操作者倾向于在邻近信号局部极小值附近冻结而非精确对齐，且邻近度本身无法预测专家 SP 质量评分。该方案通过传感器式连续距离检测补充现有技术，迈向产前扫描中的图像导航支持。