Adaptive soft erasure with edge self-attention for weakly supervised semantic segmentation: Thyroid ultrasound image...

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence

Background:

• Weakly supervised segmentation reduces the need for pixel-level annotations in medical imaging.
• Existing methods struggle with variations in nodule size and class activation map limitations, causing segmentation inaccuracies.
• Accurate segmentation of thyroid nodules is crucial for computer-aided diagnosis.

Purpose of the Study:

• To develop an improved weakly supervised segmentation method for thyroid ultrasound images.
• To overcome under- and over-segmentation issues common in current methods.
• To enhance the accuracy and quality of nodule segmentation masks.

Main Methods:

• Proposed a novel weakly supervised segmentation neural network incorporating a dual branch soft erase module.
• Implemented a scale feature adaptation module to improve sensitivity to nodule scale variations.
• Introduced an edge-based attention mechanism to enhance nodule edge segmentation.

Main Results:

• The new approach significantly outperformed existing weakly supervised semantic segmentation methods on a thyroid ultrasound dataset.
• Achieved 5.9% higher Jaccard coefficient and 6.3% higher Dice coefficient compared to baseline methods.
• Demonstrated improved segmentation accuracy and generation of high-quality segmentation masks.

Conclusions:

• The proposed method effectively improves weakly supervised segmentation for thyroid nodules.
• The combination of dual branch soft erase, scale feature adaptation, and edge attention enhances segmentation performance.
• This advancement holds promise for improving computer-aided diagnosis in thyroid ultrasound imaging.