Deep learning for improving PET/CT attenuation correction by elastic registration of anatomical data
Joshua Schaefferkoetter1, Vijay Shah2, Charles Hayden2
1Siemens Medical Solutions USA, Inc., 810 Innovation Drive, Knoxville, TN, 37932, USA. joshua.schaefferkoetter@siemens-healthineers.com.
Summary
This study introduces a deep learning method to align PET/CT scans, significantly reducing motion artifacts and improving image quality for better PET attenuation correction (AC). The technique enhances diagnostic accuracy in whole-body and cardiac imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiochemistry
Background:
- Positron Emission Tomography/Computed Tomography (PET/CT) relies on CT for attenuation correction (AC) of PET emission data.
- Subject motion between PET and CT scans introduces artifacts, compromising PET image reconstruction quality.
- Accurate CT-to-PET registration is crucial for mitigating these artifacts and improving diagnostic accuracy.
Purpose of the Study:
- To develop and validate a deep learning-based elastic registration technique for PET/CT images.
- To enhance PET attenuation correction (AC) by improving the spatial alignment between PET and CT data.
- To address artifacts caused by respiratory and voluntary motion in whole-body (WB) and cardiac myocardial perfusion imaging (MPI).
Main Methods:
- A convolutional neural network (CNN) with feature extraction and displacement vector field (DVF) regression modules was developed.
- The CNN was trained using simulated motion to predict 3D motion fields for elastic CT image warping.
- The method was evaluated on WB clinical data for motion artifact reduction and on cardiac MPI for AC improvement.
Main Results:
- The deep learning approach achieved state-of-the-art performance in PET/CT registration, significantly reducing simulated motion effects.
- Elasticly registering CT to PET distributions reduced attenuation correction artifacts in subjects with actual motion, notably improving liver uniformity.
- The technique demonstrated efficacy in improving PET AC for cardiac MPI, enhancing myocardial activity quantification and potentially reducing diagnostic errors.
Conclusions:
- Deep learning enables feasible and effective registration of anatomical (CT) to functional (PET) images for improved PET/CT attenuation correction.
- The method successfully mitigates common respiratory artifacts near the lung/liver border and misalignment from gross voluntary motion.
- This technique offers significant advantages for cardiac PET imaging, improving quantification accuracy and reducing potential diagnostic errors.
