Related Experiment Video
Updated: Sep 3, 2025
02:09
Multi-modal Pulmonary Imaging: Using Complementary Information from CT and Hyperpolarized 129Xe MRI to Evaluate Lung Structure-Function
Published on: April 12, 2024
673
Deformable CT image registration via a dual feasible neural network
Yang Lei1, Yabo Fu1, Zhen Tian1
1Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, Georgia, USA.
Medical Physics
|July 23, 2022
Summary
A deep learning (DL) method accurately registers planning CT (pCT) and quality assurance CT (QA CT) scans for cancer radiotherapy. This deformable image registration (DIR) supports contour propagation and dose evaluation, improving treatment assessment.
Area of Science:
- Medical Imaging
- Radiotherapy Physics
- Artificial Intelligence in Medicine
Background:
- Quality assurance (QA) CT scans are crucial in radiotherapy to detect anatomical changes necessitating treatment replanning.
- Accurate deformable image registration (DIR) is essential for propagating contours from planning CT (pCT) to QA CT for dose volume histogram (DVH) review.
- Deformation maps from DIR track anatomical variations and calculate accumulated dose throughout treatment.
Purpose of the Study:
- To develop a deep learning (DL)-based method for automatic DIR of pCT and QA CT scans.
- To implement a dual-feasible framework with a mutual network predicting forward and backward deformation vector fields (DVFs) simultaneously.
- To introduce a novel dual-feasible loss function for DVF regularization, preserving topology and minimizing folding.
Main Methods:
- A DL-based dual-feasible framework utilizing a mutual network for simultaneous forward and backward registration.
- Training the network with a novel dual-feasible loss function for enhanced DVF regularization.
- Experimental validation on 65 head-and-neck cancer patients (228 CTs), evaluating metrics like MAE, PSNR, SSIM, and TRE.
Main Results:
- The proposed DL method significantly improved registration accuracy, reducing mean Absolute Error (MAE) from 122.7 HU to 40.6 HU.
- Post-registration, Peak Signal-to-Noise Ratio (PSNR) increased to 30.8 dB, Structural Similarity Index (SSIM) to 0.94, and Target Registration Error (TRE) decreased to 2.0 mm.
- These quantitative improvements demonstrate the method's efficacy in pCT and QA CT DIR.
Conclusions:
- A DL-based automatic DIR method was successfully developed to align pCT and QA CT scans.
- This DIR approach enhances the current workflow for DVH evaluation on QA CTs.
- The method holds potential for facilitating studies in treatment response assessment and radiomics requiring accurate longitudinal image analysis.
