Topological data analysis visualization for interpretable assessment of AI contouring quality

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Artificial intelligence (AI) auto-segmentation tools are increasingly available.
• Conventional metrics like Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD95) do not capture regional segmentation differences.
• Interpreting contouring variations requires methods sensitive to localized discrepancies.

Purpose of the Study:

• Develop and evaluate a novel distance metric using topological data analysis (TDA) for comparing AI segmentations with reference contours.
• Assess the metric's ability to identify regional segmentation differences in organs-at-risk (OARs) and lung gross tumor volume (GTV).
• Quantify the efficiency and effectiveness of the TDA metric in highlighting contouring mismatches.

Main Methods:

• Converted segmentations into 3D point clouds.
• Clustered point clouds using K-means and the Elbow method.
• Constructed directed graphs and computed distances using unbalanced optimal mass transport, alongside DSC and HD95 calculations.

Main Results:

• The TDA metric successfully identified local regions of high deviation in OARs and GTVs.
• TDA results showed positive correlation with HD95 and negative correlation with DSC.
• The TDA computation was efficient, averaging 1.72 seconds per analysis.

Conclusions:

• A novel TDA-based metric was developed for comparing auto-segmentation of GTV and OARs.
• The TDA metric provides quantitative visualization of mismatching regions, aiding contour editing and evaluation.
• This approach offers a more nuanced assessment of segmentation accuracy than conventional metrics.