Predicting the effort required to manually mend auto-segmentations.

Area of Science:

• Medical Image Analysis
• Radiology
• Radiation Oncology

Background:

• Auto-segmentation is crucial for medical image analysis, impacting radiology and oncology efficiency.
• Current metrics like Dice Coefficient (DC) and Hausdorff Distance (HD) may not reflect clinical manual correction effort.
• Accurate evaluation metrics are needed to guide the development of superior auto-segmentation techniques.

Purpose of the Study:

• To investigate segmentation metrics that correlate with clinical manual correction effort.
• To compare established metrics (DC, HD, surDC, APL) with a new hybrid metric (Mendability Index - MI).
• To explore deep learning for predicting manual segmentation correction time.

Main Methods:

• Recorded expert mending time to quantify manual correction effort.
• Performed correlation and regression analyses on five metrics: DC, HD, surface DC (surDC), added path length (APL), and MI.
• Trained deep learning models using segmentation masks and original images to predict mending effort.

Main Results:

• The Mendability Index (MI) best indicated mending effort for sparse objects.
• Hausdorff Distance (HD) was most effective for non-sparse objects.
• Deep learning models accurately predicted mending effort, even without ground truth data.

Conclusions:

• The Mendability Index (MI) and Hausdorff Distance (HD) show promise for evaluating auto-segmentation based on object sparsity.
• Deep learning offers a novel, efficient method for assessing and enhancing auto-segmentation techniques in clinical practice.