Related Experiment Video
Updated: May 7, 2026
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
An Investigation on Cross-Tracer Generalizability of Deep Learning-based PET Attenuation Correction
Jun Hou1, Tianqi Chen1, Yinchi Zhou1
1J. Hou and T. Chen are with the Department of Biomedical Engineering, Yale University, New Haven, CT, 06511, USA. Y. Zhou, X. Chen, H. Xie, Q. Liu, and M. Xia are with the Department of Biomedical Engineering, Yale University, New Haven, CT, 06511, USA. V. Y. Panin is with Siemens Medical Solutions USA Inc, Knoxville, TN, USA. T. Toyonaga is with the Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT, 06511, USA. C. Liu is with the Department of Biomedical Engineering and the Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT, 06511, USA. B. Zhou is with the Department of Radiology, Northwestern University, Chicago, IL, 60611, USA, and the Department of Biomedical Engineering, Yale University, New Haven, CT, 06511, USA.
Deep learning models for PET attenuation correction can generalize across different radioactive tracers. A model trained on 18F-FDG PET data performed well on other tracers, reducing the need for tracer-specific training.
Area of Science:
- Medical Imaging
- Radiochemistry
- Artificial Intelligence
Background:
- Accurate attenuation correction (AC) is crucial for quantitative PET imaging.
- CT-based AC involves radiation exposure; deep learning (DL) offers a dose-free alternative by generating CT-equivalent attenuation maps (μ-CT) from PET data.
- Existing DL methods typically require tracer-specific training data, limiting their application for less common tracers.
Purpose of the Study:
- To investigate the cross-tracer generalizability of DL models for generating attenuation maps (μ-DL).
- To determine if a DL model trained on one tracer (e.g., 18F-FDG) can be effectively applied to other tracers (e.g., 68Ga-DOTATE, 18F-Fluciclovine) and vice versa.
- To evaluate the impact of different input features (μ-MLAA, λ-MLAA) on μ-DL generation performance.
Main Methods:
- Generated μ-DL using MLAA reconstruction with combined μ-MLAA and λ-MLAA inputs.
- Evaluated the performance of DL models trained on 18F-FDG when applied to 68Ga-DOTATE and 18F-Fluciclovine, and vice versa.
- Assessed the contribution of μ-MLAA and λ-MLAA as inputs for the DL network.
Main Results:
- DL models trained on one tracer demonstrated competitive performance when adapted to other tracers.
- The 18F-FDG-trained model showed the best generalizability for less common tracers with limited data.
- Combining both μ-MLAA and λ-MLAA inputs yielded the best performance, with μ-MLAA being more influential.
Conclusions:
- Cross-tracer generalizability of DL-based attenuation map generation is feasible.
- 18F-FDG-trained models offer a promising approach for dose-free AC with various PET tracers.
- Optimizing DL input features, particularly μ-MLAA, enhances AC accuracy.
