Annotation Cost Minimization for Ultrasound Image Segmentation Using Cross-Domain Transfer Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning enhances diagnostic accuracy but requires extensive annotated datasets, incurring high costs.
• Acquiring large-scale annotated medical data is time-consuming and requires expert knowledge.
• Minimizing annotation costs is crucial for broader deep learning implementation in medical imaging.

Purpose of the Study:

• To present a novel framework, SegMix, for efficient deep learning in ultrasound image segmentation.
• To enable accurate segmentation using a minimal number of manually annotated samples.
• To significantly reduce the annotation cost associated with medical image analysis.

Main Methods:

• Developed SegMix, a framework utilizing a segment-paste-blend technique to generate synthetic annotated data.
• Incorporated ultrasound-specific augmentation strategies based on image enhancement algorithms.
• Validated the framework on left ventricle (LV) and fetal head (FH) segmentation tasks.

Main Results:

• Achieved Dice/JI scores of 82.61%/83.92% for LV and 88.42%/89.27% for FH segmentation with only 10 annotated images.
• Demonstrated over 98% annotation cost reduction compared to training with full datasets.
• Maintained comparable segmentation performance despite using limited annotated data.

Conclusions:

• The SegMix framework enables satisfactory deep learning performance with very limited annotated samples.
• SegMix offers a reliable solution for reducing annotation costs in medical image analysis.
• The approach facilitates the deployment of deep learning in resource-constrained medical imaging scenarios.