RFS-UNet: Decoder-Conditioned High-Resolution Skip Recalibration for Bone-Selective DRR Synthesis

1. INTRODUCTION

• Bone-selective DRR synthesis seeks to retain skeletal structure while suppressing overlying soft-tissue contribution, without losing fine projected anatomy.Full-tissue and bone-selective projections rendered from the same CT volume under identical geometry provide paired supervision.Source: §1 · p. 1
• U-Net concatenates spatially detailed encoder features with decoder features, transferring skips independently of the decoder state.The section motivates using both encoder features and what the decoder has assembled for more adaptive transfer.Source: §1 · p. 1
• RFS uses pooled encoder and decoder features to make bounded channel-wise adjustments around identity at the two highest-resolution skips.The gate can amplify or attenuate channels without changing their spatial layout or the U-Net hierarchy.Source: §1 · p. 1
• The experiments compare decoder conditioning with encoder-only Self-RFS and use matched GPU profiling; evaluation uses a rebuilt cohort of 1,000 patient identities after an accession-split audit.The audit identified repeated patients in the original accession-level split.Source: §1 · p. 1

2. RELATED WORK

• CT-derived supervision has supported structure enhancement, bone suppression, and tissue decomposition, while DeepDRR and DiffDRR enable controlled projection pairs.This paper uses that projection-pair setting to study high-resolution feature reuse.Source: §2 · p. 1
• U-Net transfers encoder features directly, whereas Restormer redesigns processing throughout the network; SE and CBAM recalibrate features, and Attention U-Net uses decoder-conditioned spatial gates [13].UCTransNet, UDTransNet, and ProSMA-UNet also explore learned skip transfer and decoder conditioning [16].Source: §2 · p. 1
• RFS applies a bounded residual channel gate at the two finest scales, with Self-RFS isolating the conditioning signal from the scaling form.The comparison uses encoder statistics alone for Self-RFS and decoder context for RFS.Source: §2 · p. 1

3.1. Paired projection synthesis

• Paired projection synthesis renders full and bone-only densities from CT under shared geometry, then maps and letterboxes the projections for learning the full-DRR-to-bone-DRR mapping.CT and its bone mask provide construction-time supervision; the synthesis network takes the full DRR as input.Source: §3.1 · p. 1

3.2. Decoder-conditioned residual scaling

• RFS recalibrates the 512² and 256² skips before concatenation, using pooled skip and aligned decoder descriptors to predict bounded residual channel scales.The 128² and 64² skips retain direct transfers; zero initialization begins at identity, and the scale range is (0.8, 1.2).Source: Figure 1 · p. 2, §3.2 · p. 2

3.3. Training objective

• The training objective combines weighted Charbonnier intensity error, gradient loss, and an MS-SSIM structural-agreement term.The gradient and MS-SSIM terms address spatial changes and structural agreement; gradient calculations use FP32 under AMP.Source: §3.3 · p. 2

4.1. Cohort and rendering

• The audited patient-unique split assigns 700/100/200 identities to training, validation, and test, keeping each patient’s native and canonical views together.The cohort was rebuilt after an audit identified repeated patients in the original accession-level split.Source: §4.1 · p. 2

4.3. What decoder conditioning contributes

• RFS improves test PSNR from U64’s 33.27±0.04 to 33.37±0.12 dB, a gain of 0.10 dB; MAE is lower in 75.5% of patients.The patient-level result is reported with p = 3.7 × 10−14.Source: Table 1 · p. 3, §4.3 · p. 3

4.2. Training and evaluation

• Training uses AdamW without augmentation; the main RFS, Self-RFS, U64, and CBAM comparisons use seeds 2026–2028 and patient-level paired MAE analysis.Paired MAE averages the two views within each patient and then averages across three seeds.Source: §4.2 · p. 3

4.3. What decoder conditioning contributes

• RFS improves PSNR over encoder-only Self-RFS in each of three seeds, yielding a mean test gain of 0.10 dB.Both variants use bounded channel scaling; Table 2 reports patient-unique test PSNR across three seeds.Source: §4.3 · p. 3, Table 2 · p. 4

4.4. Accuracy and inference cost

• RFS runs in 4.312 ms versus CBAM’s 9.715 ms, a 2.25× speed advantage, while CBAM provides the higher-accuracy operating point at 33.48 ± 0.13 dB.The comparison uses matched RTX 5090 profiling; RFS adds latency and memory use relative to U64 but remains faster than CBAM.Source: §4.4 · p. 4, Table 3 · p. 4
• Evaluation covers one cohort and GPU using display-domain DRRs; transfer to clinical radiographs and diagnostic tasks remains open.The earlier accession-level split showed a +0.31 dB gain, while patient-unique retraining gave +0.10 dB; cohort and partition also changed, so the difference cannot isolate leakage’s effect.Source: §4.4 · p. 4

• RFS-UNet makes high-resolution skip transfer responsive to the decoder through bounded residual channel scaling.Source: §5 · p. 4
• The encoder-only control indicates that decoder context contributes to reconstruction beyond the scaling operation itself.Source: §5 · p. 4
• The evidence supports a compact skip design that improves the U-Net baseline and offers an alternative to CBAM when inference latency is a priority.Source: §5 · p. 4