# Impact of Deep-Learning-Based Respiratory Motion Correction on [18F] FDG PET/CT Test–Retest Reliability and Consistency of Tumor Quantification in Patients with Lung Cancer

## TL;DR

This study shows that using deep learning to correct for breathing motion improves the accuracy and consistency of PET/CT scans in lung cancer patients.

## Contribution

A deep-learning-based method for respiratory motion correction is shown to enhance PET/CT test–retest reliability in lung cancer.

## Key findings

- RMC reduced liver motion artifacts and improved lesion clarity and PET-CT co-registration.
- RMC led to smaller differences in tumor quantification metrics between scans compared to NMC.
- RMC increased ICCs for SUVmax, SUVmean, MTV, and TLG, indicating better TRT reliability.

## Abstract

Objectives: Respiratory motion degrades the quantitative accuracy and test–retest (TRT) reliability of fluorine-18 fluorodeoxyglucose ([18F] FDG) positron emission tomography (PET)/computed tomography (CT) in lung cancer. This study investigated whether a deep-learning-based respiratory motion correction (RMC) method improves the TRT reliability and image quality of [18F] FDG PET tumor quantification compared with non-motion-corrected (NMC) reconstructions. Methods: Thirty-one patients with primary lung cancer underwent three PET acquisitions: whole body free breathing (Scan1), thoracic free breathing (Scan2), and thoracic controlled breathing (ScanCB). Each dataset was reconstructed with and without RMC. Visual assessments of liver motion artifacts, lesion clarity, and PET-CT co-registration were scored. Lung tumors were segmented to derive standardized uptake value max (SUVmax), SUVmean, metabolic tumor volume (MTV), PET-derived lesion length (PLL), and total lesion glycolysis (TLG). Visual image scores and TRT reliability of tumor quantification were compared using Kruskal–Wallis one-way analysis of variance and intraclass correlation coefficients (ICCs). Results: RMC reconstructions achieved higher visual scores of lesion clarity and PET-CT co-registration across all lung lobes and significantly reduced liver motion artifacts compared with NMC reconstructions. Differences in SUVmax, SUVmean, PLL, MTV, and TLG between Scan2 and ScanCB were significantly smaller with RMC than with NMC. ICCs for SUVmax, SUVmean, MTV, and TLG were higher between scans with RMC than NMC reconstructions, indicating improved TRT reliability. Conclusions: The deep-learning-based RMC method improved the image quality and TRT reproducibility of [18F] FDG PET/CT quantification in lung cancer, supporting its potential for routine adoption in therapy-response assessments.