TotalRegistrator

Deformable image registration — warping one CT scan so its anatomy lines up with another — underpins longitudinal tumor assessment, multiphase contrast studies, radiotherapy planning, and image-guided intervention. Most learning-based registration networks are built for one organ at a time: a lung model, a liver model, an abdomen model, each with its own training data and hyperparameters. That fragmentation is awkward in oncology, where a single follow-up study may span thorax, abdomen, and pelvis, and it forces institutions to maintain a fleet of specialist models.

TotalRegistrator, developed at Radboud University Medical Center with collaborators in Germany and Vietnam, is a general-purpose CT registration model that aligns many anatomical regions simultaneously from a single checkpoint. Rather than scaling up architecture, it introduces a field decomposition strategy: separate UNet blocks estimate deformations for groups of spatially related structures — thorax, abdomen, bones — and a final whole-body block integrates them into one coherent displacement field, with an affine block handling initial alignment. Decomposing the problem this way keeps each block's deformation field simple enough that the whole system trains on an 11 GB consumer GPU.

The model extends a preliminary version presented at SPIE Medical Imaging 2025 by adding a dedicated bone registration block. It is positioned against uniGradICON, the first foundation model for medical image registration, and against the classical iterative solver Elastix, and is evaluated with fixed weights on three external datasets it was never trained on.

#Key Features

• Region-specific field decomposition: Independent registration blocks handle thoracic, abdominal, and skeletal structures, then compose into a single whole-body deformation field, reducing the complexity any one network must model.
• Single model across anatomy: One pipeline covers thorax, abdomen, and pelvis for both intra-patient longitudinal and inter-patient registration, replacing a collection of organ-specific networks.
• Lightweight by design: Training uses a standard UNet and fits in 11 GB of GPU memory, and full-pipeline inference on downsized volumes takes under two seconds — deliberately reproducible on hardware clinical groups already have.
• Out-of-distribution generalization: Applied without fine-tuning, the model remains competitive with specialist single-organ methods on external thoracic, abdominal, and multiphase liver data.
• Smooth deformations: A refinement stage on the whole-body block suppresses folding artifacts introduced by composing regional fields, yielding an average folding rate of 0.2%.

#Technical Details

Training uses a curated in-house dataset of 695 longitudinal intra-patient thorax-abdomen-pelvis CT pairs (591 training pairs, plus thick-section and thin-section test sets of 80 and 24 pairs), acquired on Canon, GE, Philips, and Siemens scanners with slice thicknesses from 0.8 mm to 4 mm and patient ages from 17 to 87. Coarse labels come from TotalSegmentator and were refined by clinical experts for 104 cases. Each block is a standard UNet trained with a mutual-information similarity loss, a Dice label loss, and a smoothness penalty; the affine block reuses the encoder to regress 12 transform parameters. Training took roughly 72 hours per block on a single RTX 3080 Ti. In-distribution, TotalRegistrator gives the highest average Dice across abdominal organs, with margins above 10 Dice points over uniGradICON on pancreas, spleen, stomach, and gallbladder, while uniGradICON remains stronger on lungs and some bone structures. On external data it reaches 36.20% Dice on Learn2Reg AbdomenCTCT — ahead of uniGradICON when that dataset is unseen by both, though uniGradICON reaches 48.3% when trained on it — and 97.40% Dice on the NLST lung set against uniGradICON's 98.59%. On an out-of-distribution organ excluded from both models' training, the prostate, it reaches 58.0% Dice versus uniGradICON's 61.5%.

#Applications

The model targets radiology and radiation oncology workflows where whole-body or multi-region CT is routine: aligning baseline and follow-up scans to measure lesion change, fusing multiphase contrast series for liver assessment, propagating segmentations across time points, and registering diagnostic to intra-operative imaging for navigation. Its modest hardware requirements and published Docker image make it practical for hospital research groups that cannot train large registration models, and the region-specific blocks can be run individually when only one anatomical area matters.

#Impact

TotalRegistrator argues that a carefully curated multi-organ dataset plus a decomposition of the deformation field can substitute for scale in registration, delivering broad anatomical coverage at a fraction of the training cost of a general-purpose foundation model. Its published checkpoints and Docker image make general multi-organ CT registration directly runnable, and the head-to-head external comparisons against uniGradICON give the field a concrete reference point for lightweight versus large-scale approaches. The constraints are explicit: lung alignment still trails specialist and foundation baselines, the released checkpoint covers 11 of the 14 annotated structures — omitting colon, small bowel, and prostate — training data comes from a single referral center, and the repository is licensed CC BY-NC 4.0, which excludes commercial use.