Medical volume segmentation by overfitting sparsely annotated data.

Area of Science:

• Medical Image Computing
• Artificial Intelligence in Healthcare

Background:

• Deep neural networks excel at semantic segmentation in medical imaging.
• Supervised methods require extensive, costly annotations, hindering 3D volume analysis.
• Existing sparse annotation methods still demand significant slice-by-slice labeling in 3D.

Purpose of the Study:

• To develop and evaluate methods for reducing manual slice annotation effort in 3D medical volumes.
• To assess the impact of reduced annotation on segmentation accuracy.
• To identify optimal combinations of slice selection and prediction techniques.

Main Methods:

• A two-step approach involving slice selection based on a similarity metric and subsequent segmentation prediction.
• Evaluation of various selector and predictor combinations on CT and MRI datasets.
• Quantitative assessment using Dice scores to measure segmentation performance.

Main Results:

• Achieved a Dice score of 0.969 on the Medical Segmentation Decathlon-heart dataset with only 20% of slices annotated.
• Demonstrated a positive trend of reduced annotation with maintained accuracy across multiple datasets.
• Identified specific selector-predictor combinations that maximize efficiency.

Conclusions:

• The proposed method significantly reduces the manual labeling burden for 3D medical volume segmentation.
• Provides expert recommendations for selector-predictor combinations tailored to specific tasks and objectives.
• Enables more efficient and cost-effective medical image analysis workflows.