ConfIC-RCA: Statistically Grounded Efficient Estimation of Segmentation Quality.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Assessing automatic image segmentation quality is vital in clinical settings but hindered by limited ground truth annotations.
• Reverse Classification Accuracy (RCA) estimates segmentation quality by training a segmenter on new predictions and evaluating against existing annotations.

Purpose of the Study:

• To introduce ConfIC-RCA, a novel method for automated segmentation quality estimation with statistical guarantees, even without ground truth.
• To address the challenge of reliable quality assessment in clinical workflows.

Main Methods:

• Developed In-Context RCA, utilizing in-context learning models and retrieval augmentation for efficient quality estimation with minimal reference data.
• Introduced Conformal RCA, extending RCA and In-Context RCA using split conformal prediction to provide statistically guaranteed prediction intervals for segmentation quality.

Main Results:

• ConfIC-RCA demonstrated robust performance and computational efficiency across 10 diverse medical imaging tasks.
• The method provides reliable quality estimation and statistical guarantees, crucial for automated quality control.

Conclusions:

• ConfIC-RCA offers a promising solution for automated quality control in clinical workflows by enabling fast and reliable segmentation assessment.
• The approach enhances the trustworthiness of AI-driven segmentation in medical practice.