Reproduce “Reverse Classification Accuracy: Predicting Segmentation Performance in the Absence of Ground Truth” on OpenResearch

About this paper

Vanya V. Valindria, Ioannis Lavdas, Wenjia Bai, Konstantinos Kamnitsas, Eric O. Aboagye, Andrea G. Rockall, Daniel Rueckert, Ben Glocker

The Reverse Classification Accuracy (RCA) framework predicts per-case medical image segmentation quality on new images without requiring ground truth annotations. The approach leverages a reverse classifier trained on the new image's segmentation, demonstrating effective quality prediction, with the Single-Atlas label propagation implementation showing strong correlation with actual Dice Similarity Coefficient values.

Key insights

• Segmentation quality can be predicted without ground truth by leveraging a 'reverse classifier' trained on the predicted segmentation of the target image and evaluated on a reference database with known ground truth.
• The 'Single-Atlas Label Propagation' method, when adapted as the reverse classifier, consistently performed best in predicting segmentation accuracy compared to Atlas Forests and modified CNNs.
• The Reverse Classification Accuracy (RCA) framework can effectively identify both high-quality segmentations and cases of segmentation failure on a per-case basis.

Summary by alphaXiv.