A Robust Breast ultrasound segmentation method under noisy annotations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate annotations are crucial for ultrasound image segmentation networks.
• Ultrasound image artifacts (attenuation, speckle, shadows) complicate annotation.
• Existing deep learning methods struggle with noisy ultrasound segmentation datasets.

Purpose of the Study:

• To develop an end-to-end network tolerant to noisy annotations in ultrasound segmentation.
• To address the challenge of inaccurate labels in medical imaging datasets.

Main Methods:

• Proposed a novel noisy annotation tolerance network (NAT-Net).
• Introduced a noise index (NI) to detect and dynamically correct noisy annotations during training.
• Developed a method that does not require auxiliary clean datasets or prior noise distribution knowledge.

Main Results:

• NAT-Net demonstrated superior performance over state-of-the-art methods on synthesized and real-world ultrasound data.
• Achieved a nearly 6% higher IoU on real-world datasets with complex noise.
• Showcased competitive results even on clean datasets.

Conclusions:

• NAT-Net reduces manual annotation effort and dependence on medical experts.
• Improves disease diagnosis efficiency through accurate tumor segmentation.
• Offers auxiliary strategies for ultrasound-based medical diagnosis systems.