Scientific Validation & Benchmarks
CT-ORG Cohort (N=215) • Phase 16Quantitative evaluation across held-out clinical test splits, SOTA deep learning comparisons, physical surface occlusion studies, literature-matched subsets, and whole-body cranial error resolution.
Macro MRE23.22 mm-55.8% vs PNet++
SDR ≤ 20mm54.9%+75.0% vs DGCNN
FLARE22 MRE21.30 mmZero-Shot Transfer
Brain Error-78.1%Phase 16 Retrained
S2A-Net Outperforms Classical & Deep Learning Baselines Across All Metrics
Evaluating classical linear regression, statistical shape models (SSM/PCA), PointNet++, DGCNN, and our proposed Target-Query Point Transformer across all 104 anatomical organs under strictly identical canonical coordinate alignment and train/test splits (N=215).
Proposed ArchitectureRank 1
S2A-Net (Ours)
Multi-scale surface tokens + 104 learned queries + cross-attention decoder.
Macro MRE23.22 mm
SDR ≤ 20mm54.9%
Point Cloud SOTARank 2
PointNet++ Regressor
Hierarchical set abstraction + multi-scale grouping + MLP head.
Macro MRE39.31 mm
SDR ≤ 20mm31.3%
Dynamic GraphRank 3
DGCNN Target Decoder
EdgeConv dynamic graph construction in feature space.
Macro MRE41.24 mm
SDR ≤ 20mm28.1%
Quantitative Baseline Comparison (All 104 Organs, Held-out Test Cohort)Macro Averaged Over 3 Random Seeds
| Model & Reference | Family | Macro MRE | Median Error | P90 Error | SDR ≤ 20mm |
|---|---|---|---|---|---|
Population AtlasZero-parameter mean organ coordinates from training cohort | Spatial Prior | 65.97 mm | 50.12 mm | 115.1 mm | 12.2% |
Linear Ridge RegressorDirect mapping from 4,096 flattened points to 104 targets | Classical Linear | 63.74 mm | 49.98 mm | 115.2 mm | 10.2% |
Statistical Shape Model (SSM/PCA)Surface PCA (64 principal components) + linear regression | Shape Model | 52.56 mm | 39.86 mm | 91.07 mm | 17.3% |
PointNet++ Direct RegressorMulti-scale set abstraction + MLP coordinate regressor (3-seed ensemble) | Deep Learning Point Cloud | 39.31 mm | 27.87 mm | 65.04 mm | 31.3% |
DGCNN Target DecoderPointNet++ encoder + dynamic graph convolutional decoder (3-seed ensemble) | Deep Learning Dynamic Graph | 41.24 mm | 29.18 mm | 71.45 mm | 28.1% |
S2A-Net (Target-Query Transformer)Multi-scale tokens + 104 learned queries + cross-attention (Ours) | Proposed Point Transformer | 23.22 mm | 18.79 mm | 40.12 mm | 54.9% |
Category 1 of 5