TotalRegistrator: Towards a Lightweight Foundation Model for CT Image Registration

  • 2025-08-06 13:50:27
  • Xuan Loc Pham, Gwendolyn Vuurberg, Marjan Doppen, Joey Roosen, Tip Stille, Thi Quynh Ha, Thuy Duong Quach, Quoc Vu Dang, Manh Ha Luu, Ewoud J. Smit, Hong Son Mai, Mattias Heinrich, Bram van Ginneken, Mathias Prokop, Alessa Hering
  • 0

Abstract

Image registration is a fundamental technique in the analysis of longitudinaland multi-phase CT images within clinical practice. However, most existingmethods are tailored for single-organ applications, limiting theirgeneralizability to other anatomical regions. This work presentsTotalRegistrator, an image registration framework capable of aligning multipleanatomical regions simultaneously using a standard UNet architecture and anovel field decomposition strategy. The model is lightweight, requiring only11GB of GPU memory for training. To train and evaluate our method, weconstructed a large-scale longitudinal dataset comprising 695 whole-body(thorax-abdomen-pelvic) paired CT scans from individual patients acquired atdifferent time points. We benchmarked TotalRegistrator against a genericclassical iterative algorithm and a recent foundation model for imageregistration. To further assess robustness and generalizability, we evaluatedour model on three external datasets: the public thoracic and abdominaldatasets from the Learn2Reg challenge, and a private multiphase abdominaldataset from a collaborating hospital. Experimental results on the in-housedataset show that the proposed approach generally surpasses baseline methods inmulti-organ abdominal registration, with a slight drop in lung alignmentperformance. On out-of-distribution datasets, it achieved competitive resultscompared to leading single-organ models, despite not being fine-tuned for thosetasks, demonstrating strong generalizability. The source code will be publiclyavailable at: https://github.com/DIAGNijmegen/oncology_image_registration.git.

 

Quick Read (beta)

loading the full paper ...