Truace
Subscribe

Comparative evaluation of VoxelMorpand conventional deformable image registration algorithms for thoracic 4D-CT in radiotherapy

Background Deformable image registration (DIR) is essential for thoracic four-dimensional computed tomography (4D-CT)-based radiotherapy applications. Recently, deep learning-based DIR methods such as VoxelMorph have been proposed; however, their performance relative to clinically used DIR algorithms remains unclear. Purpose This study aimed to evaluate the DIR accuracy of VoxelMorph for thoracic 4D-CT and to compare it with conventional clinical and research-oriented DIR methods. Materials and methods Thoracic…

By the Truvace desk1 min read1 sourcePeer-reviewedTRV-2026-1237· cite & verify
Comparative evaluation of VoxelMorpand conventional deformable image registration algorithms for thoracic 4D-CT in radiotherapy
Metastatic prostatic adenocarcinoma - CT scan - Case 257 (8519047084) by Yale Rosen from USA. CC BY-SA 2.0 · https://creativecommons.org/licenses/by-sa/2.0

In brief

Researchers retrospectively evaluated VoxelMorph deep learning deformable registration on thoracic 4D-CT from 64 lung cancer patients, training on 50 cases and testing on 10 against Demons, modified Demons, and ANACONDA using end-inhalation to end-exhalation registration.

By the October 2026 publication date, the test results showed VoxelMorph achieving comparable accuracy to clinical tools with high lung overlap and reduced processing time in this implementation, but the authors noted that broader standardized validation is still needed to support routine radiotherapy use.

Main points

  1. Retrospective analysis used thoracic 4D-CT from 64 lung cancer patients, with end-inhalation and end-exhalation phases for registration.
  2. VoxelMorph was trained on 50 cases, validated on 4, and tested on 10, compared against Demons (SimpleITK), modified Demons (Eclipse), and ANACONDA (RayStation).
  3. Accuracy metrics were MAE of CT values within body region, and DSC and HD95 within lung label, with median HD95 of 2.5 mm for VoxelMorph versus 2.0 mm for ANACONDA.

The gain

VoxelMorph achieved comparable deformable registration accuracy to clinical algorithms on thoracic 4D-CT, reaching 0.98 median lung Dice and 36.83 HU median MAE with shorter processing times in the evaluated implementation.

The rundown

The study used end-inhalation and end-exhalation images from 64 lung cancer patients, splitting 50 for training VoxelMorph, 4 for validation, and 10 for testing against three conventional methods.

On the 10-case test set, median MAE fell from 67.56 HU pre-registration to 36.83 HU with VoxelMorph, compared to 36.18 HU for ANACONDA and 37.39 HU for Demons, while median DSC rose to 0.98 and median HD95 was 2.5 mm.

Sources

  1. Peer-reviewedJournal of Applied Clinical Medical Physics2026-10-01

The debate