An Abdominal CT Atlas for Radiotherapy Auto-Contouring: Standardization, Uncertainty, and Quality Control
Abstract
Purpose
To create a large, quality-controlled abdominal CT atlas that enables radiotherapy auto-contouring research by providing standardized, voxel-wise annotations across diverse institutions and by supporting uncertainty-aware expert review and benchmarking.
Methods
We present AbdomenAtlas, comprising 20,460 3D CT volumes from 112 hospitals in 19 countries, with 673K high-quality structure masks. Radiologists manually annotated 22 structures in 5,246 CT volumes to establish a high-fidelity reference set, then scaled labeling to the remaining 15,214 volumes using a semi-automatic human–AI cycle. To reduce architecture bias, three segmentation backbones (Swin UNETR, U-Net, nnU-Net) generated initial labels. An attention map combined model inconsistency (cross-model variance), uncertainty (entropy), and overlap (multi-organ conflicts) to prioritize review. Six junior radiologists revised the top 5% highest-attention cases per structure under supervision, iteratively fine-tuning models; four senior radiologists performed final verification prior to release. This pipeline targets hard-to-segment structures (e.g., bowel and vessels) and standardizes contours using explicit anatomical rules.
Results
AbdomenAtlas provides per-voxel labels for 25 structures per scan and preserves original CT resolution. Manual-only annotation of the 15,214-volume remainder would require an estimated 182.9 years of radiologist time (25 structures × ~60 min/structure/CT), whereas the proposed human–AI workflow achieved a 168× acceleration while maintaining expert-level quality control. The dataset also supports supervised pretraining: a SuPreM model matched the transfer performance of a self-supervised baseline using 21 CT volumes, 672 masks, and 40 GPU-hours versus 5,050 CT volumes and 1,152 GPU-hours, and reduced downstream fine-tuning annotation needs by ~50% (e.g., comparable transfer with 512 vs 1,024 labeled cases).
Conclusion
AbdomenAtlas delivers multicenter, standardized voxel-wise abdominal CT annotations with an uncertainty-guided QA pipeline, enabling robust development and benchmarking of auto-contouring models for radiotherapy planning and related imaging applications.