PET/CT services face pressure to maintain image quality while reducing scan time and radiopharmaceutical exposure in oncology imaging. Deep learning offers a way to reconstruct clearer low-count images from shortened acquisition and lower-dose conditions. A recent analysis published in BMC Medical Informatics and Decision Making evaluated a residual U-Net model in 322 patients undergoing 18F-FDG PET/CT at The First Affiliated Hospital of Guilin Medical University. The model improved image clarity, reader agreement and lesion visibility while preserving quantitative consistency with standard 90-second images.
Shorter Scans and Lower Dose
The imaging protocol used two PET/CT systems and covered a mixed oncology population, including lung, liver, prostate, breast, thyroid, stomach and colorectal cancers, lymphoma and other malignancies. Patients fasted for more than six hours before examination and received 18F-FDG according to body weight. Imaging covered the area from the skull to mid-thigh, using the same anatomical coverage across standard and reduced-count conditions.
The standard PET acquisition used 90 seconds per bed position. Shorter image sets came from the original scan data at 30, 45, 60 and 75 seconds, while half-dose images came from down-sampling detected events before reconstruction. This approach kept reconstruction settings consistent across the different image groups.
The deep learning framework used one standard image set and five low-count image sets for each patient. A residual U-Net model worked on axial PET slices and paired each low-count image with the corresponding standard image from the same patient. Patient-level separation kept training, validation and testing groups distinct, reducing the risk of overlap between image sets from the same person.
Image Quality Gains
Deep learning processing improved visual image quality in most reduced-count groups. The strongest changes appeared in half-dose images and the shortest acquisition groups, where baseline image noise and lesion contrast were more affected. The half-dose group moved from a moderate quality score to a higher post-processing score, while 30-second and 45-second images also showed clear improvement after processing. The 75-second group changed less, consistent with its higher starting image quality.
Reader agreement also improved after processing. Agreement gains appeared in the half-dose, 30-second, 45-second and 60-second groups, suggesting more consistent image interpretation when low-count images became clearer. The change remained modest in the 75-second group, where the starting quality was already closer to the standard protocol.
Must Read: Hybrid AI Cuts Radiology Labelling Footprint
Quantitative measures followed the same direction. Structural similarity and signal-to-noise measures improved after processing in half-dose and shorter acquisition settings. SUVmax values in processed images remained close to the 90-second reference images, with no significant systematic differences. Performance also remained comparable across the two scanner systems used in the cohort, with similar quantitative image measures after processing.
Lesion Visibility and Patient Factors
Lesion visibility improved most clearly for small lesions. Detection gains were strongest in the 30-second, 45-second and 60-second acquisition groups, with the largest improvement reported in the shortest acquisition condition. Half-dose images also showed improved small lesion detection after processing. The 75-second group showed a smaller but still visible gain for small lesions.
Large lesions showed only limited change. Their starting detection rates were already high in reduced-count images, leaving less room for improvement after processing. In the half-dose group, large lesion detection rose only slightly, while the shorter acquisition groups showed similarly modest increases. This pattern suggests that the model’s main benefit appears under more challenging image conditions.
Body weight also influenced the results. Overweight patients showed larger improvements in image quality measures than normal-weight patients across several reduced-count settings. Additional voxel-level analysis showed stronger model adjustments in small lesions and in overweight patients. These subgroup findings remained exploratory. The data support further testing of deep learning approaches in patient groups where reduced counts, lesion size or body habitus make PET/CT image interpretation more difficult.
Residual U-Net processing improved low-count 18F-FDG PET/CT images created from shortened acquisition times and simulated half-dose conditions. Image quality, reader agreement and quantitative image measures improved across several reduced-count settings, while SUVmax values remained close to standard reference images. Small lesion visibility showed the clearest gain, especially under shorter acquisition conditions. The approach may support PET/CT protocols that reduce acquisition time and radiopharmaceutical use while maintaining image quality and quantitative consistency.
Source: BMC Medical Informatics & Decision Making
Image Credit: iStock
References:
Zeng Y, Chong W, Ge Z et al. (2026) Optimizing acquisition time and injected dose in 18F-FDG PET/CT imaging using deep learning: enhancing image protocol efficiency and safety. BMC Med Inform Decis Mak. https://doi.org/10.1186/s12911-026-03612-z
