BLUE RIBBON POSTER MULTI-DISCIPLINARY: Mini-Expert: Bottleneck Augmentation Reduces Boundary Outliers for Robust and Interpretable Multi-Organ Auto-Segmentation
Abstract
Purpose
Accurate multi-organ auto-segmentation is essential for efficient clinical workflows. Although nnU-Net-based models like TotalSegmentator achieve strong baseline performance, residual errors can lead to boundary outliers that require time-consuming manual corrections, particularly in small structures. We propose a Mini-Expert bottleneck augmentation to improve segmentation robustness and provide routing-derived interpretability to support clinical validation.
Methods
We integrated a lightweight Mini-Expert module at the bottleneck of the TotalSegmentator to improve robustness without modifying the overall network topology. Clinically disruptive errors in multi-organ segmentation often manifest as boundary outliers (large HD95) or isolated false-positive regions. To address this, the Mini-Expert module provides a learned “expert consultation” step that adaptively allocates additional modeling capacity to uncertain or high-risk regions. This targeted refinement aims to correct small, localized mistakes before they become large surface deviations in the final segmentation. It also providesattention maps that support qualitative interpretability. We trained both the baseline and Mini-Expert models on 1,539 CT images and evaluated segmentation across 117 anatomical structures on 6,214 organ instances using the Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95). In addition, a focused subset of 24 clinically important organs was analyzed using paired Wilcoxon signed-rank tests.
Results
Mini-Expert preserved high overlap while improving robustness, increasing mean DSC from 0.9128 to 0.9151 (p=0.002) and reducing mean HD95 from 5.94 to 4.48 mm (p=0.0001). The primary effect was suppression of extreme HD95 outliers, consistent with fewer clinically disruptive boundary failures. In the 24-organ clinical subset, Mini-Expert significantly reduced HD95 for the heart, upper lung lobe, urinary bladder, femurs, duodenum, and lumbar vertebra.
Conclusion
Mini-Expert bottleneck routing improves segmentation robustness by reducing HD95 outliers while preserving Dice. This plug-in design improves contour stability for clinically important organs. By suppressing boundary failures, it can reduce manual contour editing time and increase clinical confidence in auto-segmentation.