Learning 3-D Ultrasound Segmentation under Extreme Label Deficiency

3-D ultrasound imaging has shown great promise in clinical diagnosis by offering comprehensive volumetric assessment of organs and anatomical structures. Automatic segmentation of these anatomical structures from 3-D ultrasound images plays a critical role in precise evaluation of the clinically relevant volumetric properties of human organs. However, such segmentation algorithms typically demand a large amount of labeled data for training and the labeling process requires considerable expertise and effort, particularly in the 3-D context. In this study, our objective was to achieve accurate and robust 3-D ultrasound segmentation under extreme label deficiency.

We proposed a label-efficient 3-D ultrasound segmentation method based on a teacher-student cross-dimensional knowledge distillation framework. A 2-D teacher network, pre-trained with unsupervised representation learning, extracted rich information from ultrasound slices, with its guidance transferred to a 3-D student segmentation network, allowing the 3-D model to learn effective volumetric features from very sparse annotations.

Experiments on multiple 3-D ultrasound datasets and anatomical structures demonstrated that the proposed method consistently surpasses existing state-of-the-art techniques in segmentation accuracy, even when trained with less than 0.5% labeled data. The improved segmentations support more precise evaluation of the clinically relevant volumetric properties of human organs.

By addressing the persistent issue of label deficiency in 3-D ultrasound data, our method paves the way for greater integration of deep learning models into health care, potentially offering more accurate and efficient diagnosis and treatment in both research and clinical settings.