Nested star-shaped objects segmentation using diameter annotations.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Current deep learning segmentation models require extensive pixel-level annotations, hindering clinical application.
• A discrepancy exists between voxelwise ground-truth for model optimization and clinical annotations (e.g., diameters, counts).

Purpose of the Study:

• To develop a deep learning approach for segmenting nested star-shaped objects using diameter annotations.
• To bridge the gap between detailed voxelwise annotations and practical clinical measurements.

Main Methods:

• Proposed a method to optimize deep learning models using diameter annotations.
• Achieved this by differentiably extracting object boundary points during training for backpropagation.
• Evaluated on segmenting carotid artery lumen and wall from multisequence MR images.

Main Results:

• Reduced annotation burden to four landmarks for diameter measurement.
• Achieved state-of-the-art weakly supervised segmentation.
• Demonstrated performance comparable to full supervision.

Conclusions:

• Diameter-annotated training is effective for segmenting nested star-shaped structures.
• This weakly supervised method significantly lowers annotation requirements in medical imaging.
• The approach shows promise for clinical adoption in medical image segmentation tasks.