Abstract
Accurate synthetic Computed Tomography (sCT) generation from magnetic resonance (MR) imaging is critical for downstream clinical applications such as radiotherapy dose calculation. Optimising multi-objective loss functions remains a key challenge in medical image translation. While the current state-of-the-art nnsyn algorithm demonstrateshighlyrobustperformanceinsCTgeneration,itsstaticallyweighted multi-objective loss function lacks the adaptability required to fully resolve gradient interference in anatomically complex regions such as the abdomen. To address this, we present a dynamic loss weighting framework utilising Random Loss Balancing with Random Lookback (ReLo- BRaLo). By dynamically balancing the multi-objective loss function as the model trains, our approach systematically avoids gradient conflict by initially prioritising global Mean Squared Error before shifting focus to Anatomical Feature Perception Loss. We evaluated the framework on 175 abdomen MR-CT patient pairs from the SynthRAD2025 Grand Challengedataset.TheReLoBRaLoimplementationyieldedstatistically significant improvements (p < 0.05) over the static nnsyn state-of-theart baseline, SoftAdapt relative balancing and Gradient Normalisation (GradNorm)implementationacrossallevaluatedmetrics.Itachievedsuperiorvoxel-wisesimilarity(MAE,PSNR,MS-SSIM)andgeometricconsistency(mDice,HD95).Furthermore,thisstructuralfidelityisachieved with a negligible 0.3% computational overhead per epoch. Our framework provides a highly scalable, mathematically principled approach to lossbalancing,successfullybridgingthegapbetweenglobalimagestructureandhigh-fidelityanatomicaldetailrequiredforrobustmedicalimage translation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SASHIMI_040.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=Ij6QG76pKx
BibTex
@InProceedings{deMic_Relative_MICCAISAT2026,
author = { de Francesco, Michael AND Xin, Bowen AND Dowling, Jason AND Hoesl, Michaela AND Wang, Xiuying},
title = { { Relative Loss Balancing for High-Fidelity Synthetic CT Generation in the Abdomen } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}
