Selective Kidney Segmentation with Rejection Option via Machine Reasoning

A machine reasoning framework with rejection option reduced false positives and anatomically implausible errors in CT kidney segmentation, while preserving acceptance of valid outputs. Robustness persisted under noise.

• nnUNet integrated with SimpleMind reasoning using FMR, AR, and LCCAPS outperformed LCCA in reducing false positives and invalid outputs.
• Candidate-level acceptance and rejection evaluation metrics were introduced.
• Internal five-fold cross-validation and robustness testing under added noise were performed.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.