Medical Instrument Segmentation in 3D US by Hybrid Constrained Semi-Supervised Learning

Area of Science:

• Medical imaging
• Deep learning
• Image-guided intervention

Background:

• Accurate medical instrument segmentation in 3D ultrasound (US) is crucial for image-guided interventions.
• Training deep neural networks for this task typically requires extensive labeled data, which is costly and time-consuming to acquire.

Purpose of the Study:

• To develop a semi-supervised learning (SSL) framework for 3D US instrument segmentation that minimizes annotation effort.
• To introduce a novel Dual-UNet architecture incorporating a hybrid loss function for effective utilization of unlabeled data.

Main Methods:

• A Dual-UNet model was proposed for instrument segmentation in 3D ultrasound.
• A hybrid loss function combining uncertainty and contextual constraints was developed to leverage unlabeled data for SSL training.
• Uncertainty constraints utilized prediction uncertainty, while contextual constraints exploited image context for improved voxel-wise estimation.

Main Results:

• The proposed SSL framework achieved Dice scores ranging from 68.6% to 69.1% across multiple datasets.
• Inference time was approximately 1 second per volume, comparable to supervised methods.
• Performance surpassed state-of-the-art SSL methods in 3D ultrasound instrument segmentation.

Conclusions:

• The developed semi-supervised learning framework effectively segments instruments in 3D ultrasound with reduced annotation requirements.
• The Dual-UNet with its hybrid loss function demonstrates superior performance and efficiency compared to existing SSL approaches.
• This method offers a practical solution for improving image-guided interventions through accurate and rapid instrument segmentation.