Deep learning-enhanced data-driven gating improves FDG PET/CT clinical image quality

Journal article

Background: Respiratory motion can affect PET image quality. One way to reduce motion effects is respiratory gating. The objective of this study, as we seek to further optimise Data-Driven Gating (DDG) algorithms, is to compare two types of DDG phase-gating methodologies: Method-1 has fixed quiescent period offset while Method-2 has an optimised offset for each cycle based on the amplitude of the waveform. The use of Deep Learning is becoming more prevalent for medical images. A previously validated Deep Learning Enhancement (DLE) algorithm will be assessed in conjunction with DDG PET data as an additional method to improve clinical images impacted by respiratory motion. Six reconstructions were assessed: Ungated 3 min (Clinical Standard), Ungated 6 min (Gold Standard), both gating methods with BSREM reconstructions, and both gating methods with OSEM + DLE. These six reconstructions were compared with data from the NEMA IQ phantom, placed on the QUASAR motion platform. Contrast Recovery (CR), Background Variability (BV) and Contrast to Noise Ratio (CNR) were calculated across 6 hot spheres. The same six reconstructions were assessed for 39 FDG PET-CT patient scans with lesions in the lungs or liver. All patients were identified as “high motion” patients with lesions in the region of interest. An experienced radiologist ranked the images and scored them on a 5-point Likert Scale for lesion detectability, diagnostic confidence, and image quality. Lesion maximum standard uptake value (SUVmax) and liver background noise were analysed across all images. Results: The DLE methods with both gating methods demonstrated significantly better CNR than the ungated images in the phantom data (p<0.05). The performance of DLE with DDG data was also supported by the clinician’s preferences and decreased liver noise across the patient reconstructions. Both phantom and patient data indicated a slight preference for Method 2 for patients with irregular breathing. Conclusions: DLE algorithm effectively produces BSREM-like images from OSEM inputs with data-driven gated data. Clinician preference strongly indicated a preference for the two DDG DLE methods (p<0.05). Preliminary results may indicate an improved image quality with cycle-specific phase offset gated images for patients with irregular breathing patterns characterized by breathing period variance.

Authors: Wilson, Z; Dedja, M; Su, K; Johnsen, R; Bradley, KM

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: SpringerOpen
• Journal: EJNMMI Physics
• Volume: 13
• Issue: 1
• Article number: 46
• Publication date: 2026-04-05
• Acceptance date: 2026-02-27
• DOI: 10.1186/s40658-026-00851-x
• EISSN: 2197-7364
• ISSN: 2197-7364
• Language: English
• Keywords: Deep learning; Gating; PET; Respiratory motion