Source-linked AI summary
Region-Weighted Losses and Model Fusion for Cross-Modal PET Attenuation Correction
Khoa Tuan Nguyen, Joris Vankerschaver, Wesley De Neve
TL;DR
BIC-MAC asks for pseudo-CT synthesis from NAC-PET, DIXON MRI, and a topogram, evaluating both the pseudo-CT and reconstructed AC-PET. The paper aligns a region-weighted loss with attenuation-coefficient scoring and fuses two models, placing first overall on public validation.
Problem
HU-space loss mismatches the CT metric’s attenuation-coefficient scoring: equal HU errors count differently above and below its 47-HU knee.
Source: §2.2 · p. 4Method
The approach uses region-weighted L1 error in Carney μ space, adds unregistered DIXON MRI as input channels, and combines two independently trained models with a fixed convex blend.
Source: §2.2 · p. 4, §3 · p. 6, §3.1 · p. 7Results
Mean rank 3.50 across 26 submissions placed the submitted fusion first overall on the 16 August 2026 public validation leaderboard.
Source: §3.2 · p. 7Takeaways & Limitations
The validation results support fusion as a way to achieve strong overall standing despite mixed metric-specific ranks.
Source: §3.2 · pp. 7–8MuNet4 was not trained with LNGF, leaving the combined effect of direct MRI input and LNGF unmeasured.
Source: §3 · pp. 6–7
Abstract
from arXiv · showhide
We describe our approach to the Big Cross-Modal Attenuation Correction (BIC-MAC) challenge, which asks for a pseudo-CT in Hounsfield Units to be synthesized from Non-Attenuation-Corrected PET (NAC-PET), DIXON MRI and a topogram, and scores both the pseudo-CT and the Attenuation-Corrected PET (AC-PET) reconstructed from it. Three ideas carried our improvements over the organizers' 3D U-Net baseline. The loss matters more than the architecture: we compute the $L_1$ error in the Carney attenuation-coefficient ($μ$) space that the CT metric itself uses, weighted by anatomical region. Only once that loss was in place did the unregistered DIXON MRI work as extra input channels. A fixed convex combination of two independently trained models then beat both of its members on three of the four metrics and ranks first overall on the public validation leaderboard.
1 Task and Evaluation
The section defines whole-body pseudo-CT synthesis from NAC-PET, two-phase DIXON MRI, and a 2D topogram, evaluated through the fixed STIR pipeline with rank-based multi-metric scoring.
1 Task and Evaluation
- The task predicts whole-body HU pseudo-CT from NAC-PET, two-phase DIXON MRI, and a 2D topogram for fixed STIR reconstruction.Evaluation uses four lower-is-better metrics: whole-body CT µ-map MAE, whole-body SUV MAE, organ bias, and a dataset-level brain-outlier score; final ranking is the mean of per-metric ranks.Source: §1 · p. 1
- Below 47 HU, the Carney HU-to-µ slope is 9.6×10^-5 cm^-1 per HU versus 5.1×10^-5 above, motivating CT evaluation in µ space.The below-knee segment is 1.88 times steeper, so the same HU error produces almost twice as much µ error in soft tissue as in bone.Source: §1 · p. 1, Figure 1 · p. 2
- Pipeline 4 mm FWHM smoothing removes 3.6% of mean µ and reduces rib-and-spine profile peak-to-peak range from 77 to 37 × 10^-3 cm^-1.Pseudo-CT detail finer than about 4 mm is invisible to the three AC-PET metrics but remains visible to the CT metric.Source: Figure 1 · p. 2, §1 · p. 2
2 Method
The method retains a residual 3D U-Net backbone while varying input channels and loss terms, then voxel-wise fuses independently trained models into one pseudo-CT volume.
2.1 Network
- MuNet adds attention-gated skip connections, bottleneck self-attention, and two auxiliary CT-prediction heads to the organizers’ residual 3D U-Net.MuNet2 reads NAC-PET and topogram; MuNet4 adds raw in- and out-phase DIXON MRI, supplied unregistered because deployment lacks ground-truth CT for registration.Source: §2.1 · pp. 3–4, §2 · p. 3
2.2 Lµ: a region-weighted loss in the metric’s own space
- Lµ measures L1 error in the Carney attenuation-coefficient space used by the CT metric, correcting HU loss’s 1.88-fold underweighting below 47 HU.Loss is restricted to the body mask, excluding face and scanner bed, and region weights emphasize scored anatomy; these weights also enabled scratch training to learn bone where plain L1 failed.Source: §2.2 · p. 4
2.3 Lgrad: matching edge strength
- Lgrad penalizes differences in predicted and target CT gradient magnitudes, emphasizing soft-tissue boundaries where error concentrates.Unlike per-voxel Lµ, it encourages edge strength in the same locations and primarily targets boundary sharpness.Source: §2.3 · p. 4
2.4 LAnPer: an Anatomical Perception loss
- 5.8% lower organ bias (3.34 to 3.14) and 27.4% lower brain-outlier score (0.0261 to 0.0189) came from adding LAnPer in a matched pair.LAnPer matches intermediate decoder features from a frozen TotalSegmentator applied to predicted and ground-truth CTs, targeting anatomical agreement; the segmenter is used only during training.Source: §2.4 · p. 5
2.5 LNGF: aligning boundaries with the MRI
- LNGF aligns MRI and predicted-CT boundary directions in soft tissue, using gradient orientation rather than magnitude because DIXON MRI has strong soft-tissue contrast but little bone signal.The MRI contributes through this loss alone, so inference still uses only NAC-PET and topogram.Source: §2.5 · p. 5
2.7 Fusion of finished models
- Finished models are combined by a fixed voxel-wise convex average in HU, with positive weights summing to one, so the container still submits one pseudo-CT.Members train independently; the reported pair and weights were selected from measured metrics and confirmed by full reconstruction, while three-member blends did not outperform the best pair.Source: §2.7 · p. 6
3 Results
The submitted fusion ranks first overall on the public validation leaderboard, although the second-placed entry beats it on three of four metrics. Results show that the fusion gain depends on the model pair, while loss choice and MRI inputs shape single-model performance.
3 Results
- All models trained on all 75 released subjects with Adam and a cosine schedule, using two patches per optimizer step; training took 18 to 60 hours.Patch size was 2083, or 448×448×48 where LLOR was active, filling one 48 GB RTX A6000.Source: §3 · p. 6
- Changing the loss reduced CT µ-MAE from 0.006610 to 0.005803; adding MRI inputs gave the best single-model CT µ-MAE, 0.005679, while LNGF traded CT performance for AC-PET metrics.Adding LLOR to either LNGF or MRI shifted results similarly; against row 5, row 6 was 0.42% worse on CT and improved SUV MAE by 11.4%, organ bias by 15.0%, and brain outlier by 53.9%. The authors attribute the LNGF CT tradeoff to its effect on fine structure, which is smoothed by the 4 mm blur before the AC-PET projection.Source: §3 · p. 6
- MuNet4 with LNGF was not tested; the authors considered the methods partly redundant because MuNet4 directly reads MRI, unlike LNGF, but left the gap unmeasured.They chose to record the gap rather than infer the combination's performance.Source: §3 · pp. 6–7, §3.1 · p. 7
- Row 9, the submitted model, improves on row 8 by 1.3% on SUV MAE, 1.9% on organ bias, and 4.9% on brain outlier, with 0.11% more CT µ-MAE.Refining row 6 left training loss at 0.0133 against 0.0132, so training loss alone would not have signaled the gain.Source: §3 · p. 6, Table 1 · p. 7
3.1 Why the fusion gain depends on the pair
- Row 8 beats both component models on CT µ-MAE, at 0.005627 against 0.005703 and 0.005958, and is 2.6% below their linear-interpolation value.The brain-outlier gain is attributed to averaging partly independent predictors, which suppresses tail errors more than it shifts means.Source: §3.1 · p. 7, Table 1 · p. 7
- The fusion gain depends on the pair: averaging rows 4 and 5 gained almost nothing on brain outlier, whereas averaging row 4 with row 6 produced row 8.Rows 4 and 5 both scored 0.0168 on brain outlier; row 6 scored 0.0078, and the authors selected the pair using measured metrics and confirmed it with a full reconstruction.Source: §3.1 · p. 7, Table 1 · p. 7
3.2 Leaderboard
- On the public validation leaderboard of 16 August 2026, fusion ranks first overall with mean rank 3.50 across 26 submissions, placing 7th on CT µ-MAE, 2nd on SUV MAE, 3rd on organ bias, and 2nd on brain outlier.The second-placed entry has a mean rank of 3.88, beats the fusion on three of four metrics, and loses only on CT µ-MAE, where the fusion is 3.3% ahead.Source: §3.2 · pp. 7–8