BoxSegUS: A Spatial Consistency Box Supervised Multi-Class Segmentation with Prior and Boundary Constraint for...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of anatomical structures in ultrasound images is crucial for clinical applications like quantitative assessment, fetal development monitoring, and disease diagnosis.
• Fully supervised methods require extensive pixel-level annotations, driving research into less data-intensive weakly supervised techniques.
• Existing box-supervised methods are often limited to single-structure segmentation and struggle with complex multi-structure ultrasound images due to variations, ambiguous cues, and overlapping annotations.

Purpose of the Study:

• To develop a robust box-supervised framework, BoxSegUS, for accurate segmentation of multiple anatomical structures in ultrasound images.
• To address challenges in ultrasound segmentation, including spatial variations, unreliable local appearance cues, and class-assignment ambiguity caused by overlapping bounding boxes.

Main Methods:

• BoxSegUS utilizes bounding-box annotations for segmentation, incorporating weak-strong spatial consistency for robustness against variations.
• A detection-prior global context modeling mechanism is employed to mitigate unreliable local appearance cues.
• Inter-/intra-class anatomical priors guide mask evolution, and a Soft Projection loss with sparse boundary regularization refines segmentation and handles imperfect annotations.

Main Results:

• BoxSegUS demonstrates superior performance compared to existing box-supervised baselines across four diverse ultrasound datasets.
• The framework achieves competitive results when compared to fully supervised segmentation models.
• Experiments included three multi-structure datasets and one external thyroid nodule dataset, validating the generalizability of BoxSegUS.

Conclusions:

• BoxSegUS offers an effective solution for multi-structure ultrasound image segmentation using only bounding-box annotations.
• The proposed methods enhance robustness and accuracy, overcoming limitations of previous box-supervised approaches in complex medical imaging scenarios.
• This framework provides a promising direction for advancing weakly supervised medical image segmentation, reducing annotation burden while maintaining high performance.