Source-linked AI summary

Region-Weighted Losses and Model Fusion for Cross-Modal PET Attenuation Correction

Khoa Tuan Nguyen, Joris Vankerschaver, Wesley De Neve

arXiv:2608.21881v1cs.CV

TL;DR

The paper addresses mismatches between training losses and attenuation-correction metrics, including how to use MRI information and improve reconstructed PET outcomes. It uses a metric-aligned, region-weighted μ loss, auxiliary network components, MRI inputs, line-of-response modeling, and model fusion; the submitted fusion ranks first overall on the public validation leaderboard.

  • Problem

    Training pseudo-CT errors in HU can misprice errors relative to the CT metric, while organ activity depends on attenuation integrated along lines crossing the body rather than only within the organ.

  • Method

    The approach retains the residual 3D U-Net backbone, adds attention gates, bottleneck self-attention, auxiliary multiscale heads, and a region-weighted L1 loss in Carney μ space inside the body mask.

  • Results

    The submitted fusion ranks first overall with mean rank 3.50 across 26 submissions, placing second on SUV MAE and brain outlier, third on organ bias, and seventh on CT μ-MAE.

  • Takeaways & Limitations

    The loss produced the largest single improvement, unregistered MRI achieved the best single-model CT μ-MAE, and fusion improved three non-CT metrics over its comparison blend.

  • Takeaways & Limitations

    MuNet4 with LNGF was never trained, leaving their potentially redundant combination unmeasured.

Abstract

from arXiv · show

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 IDLab, ELIS, Ghent University, Belgium

The paper is affiliated with Ghent University entities in Belgium and Korea, and lists model fusion and PET attenuation correction among its keywords.

  • The authors are affiliated with IDLab and ELIS at Ghent University in Belgium.
  • The listed keywords include model fusion, PET attenuation correction, pseudo-CT synthesis, and region-weighted loss.

1 Task and Evaluation

The task predicts whole-body pseudo-CT from NAC-PET, DIXON MRI, and a topogram, then evaluates both CT attenuation accuracy and reconstructed PET quality through a fixed pipeline.

  • The input comprises NAC-PET, two-phase DIXON MRI, and a 2D topogram for whole-body pseudo-CT synthesis in HU.
  • The CT µ-map MAE converts HU to linear attenuation coefficients using the Carney bilinear model before scoring.
  • The HU-to-µ map has a knee near 47 HU, making equal HU errors worth 1.88 times more µ below than above the knee.
  • A 4 mm FWHM Gaussian smooths the µ-map before forward projection, hiding detail finer than about 4 mm from the three AC-PET metrics while CT still sees it.

2 Method

The method combines an augmented residual 3D U-Net, metric-aligned region-weighted losses, MRI-derived signals, line-of-response supervision, and voxel-wise fusion of independently trained models.

  • 2.1 Network: MuNet augments the residual 3D U-Net baseline with attention-gated skips, bottleneck self-attention, and auxiliary half- and quarter-resolution heads.
  • 2.1 Network: MuNet2 reads NAC-PET and the topogram, whereas MuNet4 adds unregistered raw in- and out-phase DIXON MRI channels.
  • 2.2 Lµ: a region-weighted loss in the metric’s own space: The core loss measures L1 error in Carney attenuation-coefficient µ space instead of normalized HU space, aligning optimization with the CT metric.
  • 2.2 Lµ: a region-weighted loss in the metric’s own space: The loss scores only voxels inside the body mask and uses anatomical region factors, while training-only ground-truth segmentations do not add inputs at deployment.
  • 2.2 Lµ: a region-weighted loss in the metric’s own space: Region weighting enables random-initialization training to learn bone, whose absence otherwise damages the skull-driven brain-outlier score.
  • 2.6 LLOR: a loss along Lines Of Response: LLOR supervises µ differences along whole-body and organ-crossing lines after 4 mm smoothing, matching the spatial frequencies visible to AC-PET metrics.
  • Finished independently trained models are averaged voxel-wise in HU using fixed positive weights summing to one before scoring.

3 Results

The results show that region-aware loss design and carefully selected model fusion improve complementary validation metrics, while an unmeasured MuNet4–LNGF combination remains a documented gap. The submitted fusion ranks first overall on the public validation leaderboard, despite a narrow margin.

  • Training setup: 18 to 60 hours were required to train each model, depending on the schedule, using all 75 released subjects.Training used Adam with a cosine schedule and two patches per optimizer step.
  • Single-model progression: 0.005679 was the best single-model CT µ-MAE, achieved when unregistered MRI was added as input channels.This result supports the usefulness of unregistered DIXON MRI as an additional input.
  • Loss and regional trade-offs: 0.42 % worse on CT but 11.4 %, 15.0 % and 53.9 % better on SUV MAE, organ bias and brain outlier resulted from adding LLOR to the MRI model.LLOR was the only tested term that improved organ bias.
  • Fusion refinement: 1.3 %, 1.9 % and 4.9 % improvements on SUV MAE, organ bias and brain outlier came from refining one blend member, for 0.11 % more CT µ-MAE.The training loss remained nearly unchanged, at 0.0133 versus 0.0132, so training loss alone would not have selected the refinement.
  • Limitations: The results leave one table gap: MuNet4 with LNGF was never trained, although the two additions are partly redundant by construction.The authors state that they expect less from this combination but did not measure it.
  • Fusion pair selection: 2.6 % below linear interpolation was the CT µ-MAE of row 8, which beat both of its members at 0.005627 versus 0.005703 and 0.005958.The gain depended on the selected pair: averaging models with similar brain failures produced little brain improvement, whereas row 8 combined complementary behavior.
  • Leaderboard: 3.50 mean rank across 26 submissions placed the submitted fusion first overall on the public validation leaderboard on 16 August 2026.Its per-metric ranks were 7th on CT µ-MAE, 2nd on SUV MAE, 3rd on organ bias and 2nd on brain outlier; the second-place entry had mean rank 3.88 and beat it on three metrics.
Loading 2608.21881v1…