Selective segmentation with rejection option: a machine reasoning approach and evaluation metrics in CT kidney segmentation.
Authors
Affiliations (1)
Affiliations (1)
- University of California, Los Angeles, Center for Computer Vision & Imaging Biomarkers, Department of Radiology, Los Angeles, California, United States.
Abstract
Deep learning models for medical image segmentation achieve high average performance but can produce anatomically implausible errors that undermine trust and necessitate manual review. This work addresses this limitation through selective segmentation with a rejection option, enabling systems to accept plausible results while rejecting questionable ones. A nnUNet model trained for kidney segmentation on the KiTS23 dataset was integrated into the SimpleMind (SM) reasoning framework. Kidney segmentation candidates were evaluated and selected using SM reasoning methods fuzzy membership reasoning (FMR), adaptive reasoning (AR), and largest connected component analysis per side (LCCAPS), and performance was compared with the commonly used largest connected component analysis (LCCA). Performance was assessed via false positive rate and false negative rate across five validation folds. Candidate-level acceptance and rejection metrics were also introduced to evaluate selective performance. Robustness was further examined under added noise. Reasoning-based approaches (FMR, AR, and LCCAPS) reduced false positives and anatomically invalid outputs more effectively than LCCA, while preserving acceptance of valid segmentation candidates. Results persisted under noise, avoiding acceptance of implausible results. Incorporating reasoning into DL segmentation supports selective acceptance of anatomically valid outputs and rejection of invalid ones, increasing reliability and autonomy in medical imaging workflows. This selective segmentation framework provides a practical path toward trusted, semi-autonomous clinical deployment.