A pilot study published in EJNMMI Physics demonstrates that a deep learning-based denoising algorithm can improve the diagnostic image quality of low-dose 18F-FDG breast PET/MRI scans. The research, conducted as a secondary analysis of a single-institution prospective study, evaluated whether reducing the radiation dose to a level comparable to digital mammography could maintain acceptable image quality for clinical use. The findings suggest that deep learning enhancement can mitigate the increased image noise typically associated with reduced radiotracer activity, potentially facilitating broader clinical utilization of PET/MRI for breast cancer assessment.
The study focused on 23 women with biopsy-proven primary invasive breast cancer, totaling 24 malignancies. The original prospective study was conducted between January 2016 and August 2018, involving prone simultaneous breast PET/MRI acquisitions. For this pilot analysis, the researchers used retrospective random undersampling of the PET data to simulate a 90% reduction in injected activity. This simulated low-dose (SLD) approach approximated an effective dose of 0.72 mSv, which is similar to the effective dose of full-field digital mammography, ranging from 0.44 to 0.49 mSv.
To generate the low-dose images, the team segmented the PET list-mode data into very short time intervals and randomly selected approximately one-tenth of the events acquired in the fully sampled study. This method was chosen over simply shortening the scan duration to avoid capturing motion and other artifacts inherent in longer acquisitions. The simulated low-dose images were then processed using a vendor-provided, deep learning-based denoising (DN) algorithm. This investigational model, trained on multi-center whole-body oncology PET/CT exams, was designed to produce images with reduced noise while preserving diagnostic features.
Three distinct PET datasets were generated for each participant for comparison: the simulated low-dose images without denoising (SLD), the simulated low-dose images with denoising (DN), and the original full-dose images. Three board-certified radiologists and nuclear medicine physicians independently reviewed the SLD and DN datasets in a blinded, randomized order. The readers evaluated the images using a 5-point scale for artifacts, perceived signal-to-noise ratio, image sharpness, and overall image quality. They also assessed diagnostic image quality and lesion conspicuity relative to the full-dose acquisition to determine clinical acceptability.
The results indicated that the denoised datasets significantly outperformed the non-denoised low-dose images in multiple quality metrics. Compared to the SLD images, the DN datasets showed fewer artifacts, improved perceived signal-to-noise ratio, better image sharpness, and higher overall image quality scores, with all differences reaching statistical significance (p < 0.001). When compared to the full-dose images, the DN datasets were rated as equivalent or slightly inferior in diagnostic image quality in 89.9% of the reads. Readers preferred the DN images over the SLD images in 76.8% of cases, whereas the SLD images were preferred in only 4.3% of cases.
Objective quantification of image noise further supported the subjective reader assessments. The study measured image noise by calculating the standard deviation of the standardized uptake value (SUV) in the liver. The mean SUV standard deviation was 0.50 for the SLD images, 0.23 for the DN images, and 0.11 for the full-dose images. The denoising process significantly reduced image noise compared to the SLD images (p < 0.0001), although the full-dose images retained the lowest noise levels. This objective reduction in noise correlated with the readers’ perception of improved signal-to-noise ratio in the DN datasets.
The study also evaluated the impact of denoising on the semi-quantification of radiotracer uptake in both tumor and normal breast tissue. Bland-Altman analyses showed excellent agreement for tumor SUVmax and SUVmean across all three datasets. For normal breast tissue, agreement was good between the DN and full-dose images for SUVmax, and excellent for SUVmean between SLD and DN images. While tumor SUVmax was slightly higher in the SLD and full-dose images compared to the DN images, there were no significant differences in tumor SUVmean among the datasets. The authors noted that SUVmax is sensitive to image noise due to its reliance on a single voxel, which explains the slight variations observed in the low-dose images.
Limitations of the study include its relatively small sample size, which was constrained by the number of participants in the original prospective study. The retrospective use of undersampled PET data serves as an approximation of actual low-dose imaging but requires further prospective validation to confirm these findings in a real-world clinical setting. Additionally, the study did not evaluate the effect of denoising on small lesions, as the original study included only tumors measuring at least 10 mm to avoid partial volume effects on SUV semi-quantification. The performance of the denoising algorithm may also vary across different PET/MRI systems, as this study utilized a single vendor’s model on a specific time-of-flight capable system.
The findings contribute to the growing body of evidence supporting the feasibility of low-dose PET/MRI for breast cancer imaging. Previous studies have explored dose reduction strategies for dedicated breast PET and positron emission mammography, but this research specifically addresses the application of deep learning to breast PET/MRI. By reducing the effective dose to a level comparable to mammography, the study addresses a key barrier to the clinical adoption of PET/MRI, which is the higher radiation exposure compared to traditional breast imaging modalities. The potential benefits include reduced radiation risk for patients, particularly in high-risk screening populations, and improved patient comfort due to potentially shorter acquisition times.
The authors conclude that acceptable image quality may be achieved with low-dose 18F-FDG breast PET/MRI using deep learning-based denoising. This approach warrants further prospective validation in larger trials that include actual reductions in injected activity and a broader range of lesion sizes. Future research should also compare the deep learning model with other reconstruction techniques and evaluate its diagnostic accuracy for cancer detection in a real-world clinical setting. If validated, this technology could extend the clinical applicability of PET/MRI, offering patients the reassurance of low radiation exposure without compromising the diagnostic integrity necessary for effective cancer management.
The deep learning enhancement model used in this investigation was not developed specifically for breast PET/MRI. Instead, the vendor-provided algorithm was trained on multi-center, multi-scanner whole-body oncology 18F-FDG PET/CT exams. The model is a three-dimensional residual convolutional encoder-decoder network, often referred to as a U-Net, which was trained through supervised learning cycles. It learned to map high-noise input images, reconstructed using conventional ordered-subset expectation maximization, to low-noise reference images reconstructed with a regularized algorithm. The goal was to produce output images that mimic the quality of the regularized reconstruction while starting from the noisier low-dose data. The researchers noted that the injected activity levels in this breast study were similar to those in the training data, and the scan duration was longer, placing the simulated low-dose conditions within the range of the model’s training distribution.
Quantitative analysis of radiotracer uptake revealed subtle differences in how the denoising algorithm affected maximum standardized uptake values. For the 24 breast tumors, SUVmax was significantly higher in the simulated low-dose images without denoising and the full-dose images compared to the denoised images. The mean SUVmax for the denoised images was 10.2, whereas it was 11.75 for the non-denoised low-dose images and 11.68 for the full-dose images. In contrast, the mean standardized uptake values for the tumors showed no significant differences among the three datasets, hovering around 4.6 to 4.8. The authors explained that SUVmax is derived from a single voxel and is therefore highly sensitive to statistical fluctuations and image noise, which accounts for the observed variations in the low-dose conditions. The mean values, which average uptake over a larger volume, remained stable across the different image processing methods.
Similar patterns emerged in the normal breast tissue measurements. The SUVmax for normal contralateral fibroglandular tissue was higher in the non-denoised low-dose images compared to both the denoised and full-dose images. However, the SUVmean for normal breast tissue did not differ significantly between the datasets, with values consistently around 1.4. Bland-Altman analyses indicated excellent agreement for tumor SUVmax and SUVmean across all three datasets. For normal breast tissue, agreement was good between the denoised and full-dose images for SUVmax, and excellent for SUVmean between the non-denoised low-dose and denoised images. These findings suggest that while the denoising process slightly alters peak uptake values, it preserves the overall quantitative integrity of the radiotracer distribution in both malignant and normal tissues.
The study’s design included specific technical choices to ensure the simulated low-dose images accurately reflected the expected noise characteristics of a reduced-dose scan. Rather than simply shortening the acquisition time by 90%, which would introduce motion artifacts and other inconsistencies, the researchers used random undersampling of the list-mode PET data. This method involved segmenting the data into very short time intervals and randomly selecting approximately one-tenth of the events to form the final sinogram. This approach better approximates the statistical noise of a lower injected activity without the confounding factors of patient motion. The full-dose images were reconstructed using a time-of-flight block sequential regularized expectation maximization algorithm, while the low-dose images used a standard ordered-subset expectation maximization reconstruction before denoising. This difference in reconstruction algorithms is consistent with how the deep learning model was originally trained, minimizing potential bias in the comparison.
Reader agreement on the image quality assessments was moderate, with a Fleiss kappa of 0.034, largely due to the strong preference for the denoised images. However, 78.3% of the cases showed agreement where two readers scored the images identically and the third differed by only one point on the five-point scale. The primary reasons for unfavorable scores in the low-dose images included artifacts such as image streaking and haze around the breast, as well as difficulty detecting smaller lesions due to image smoothing. These observations highlight that while denoising significantly improves overall image quality, it does not entirely eliminate the challenges associated with low-dose imaging. The study’s focus on image quality preference rather than diagnostic accuracy for cancer detection is a key limitation, as the ultimate clinical value of low-dose PET/MRI depends on its ability to reliably detect and characterize lesions in a broader patient population.
Subject of Research: Deep learning enhancement of low-dose 18F-FDG breast PET/MRI image quality
Article Title: Image optimization for low-dose 18F-FDG breast PET/MRI using deep learning: a pilot study
Article References: Keigley, Q. J., Henze Bancroft, L. C., Salem, K., Su, K.-H., Mehranian, A., Spangler-Bickell, M. G., McMillan, A. B., Perlman, S. B., Strigel, R. M., Bradshaw, T. J., Hurley, S. A., Mao, L., & Fowler, A. M. (2026). Image optimization for low-dose 18F-FDG breast PET/MRI using deep learning: a pilot study. EJNMMI Physics, 13(1), Article 100. https://doi.org/10.1186/s40658-026-00918-9
Image Credits: AI Generated
DOI: 10.1186/s40658-026-00918-9
Keywords: breast cancer, PET/MRI, deep learning, radiation dose, image denoising, 18F-FDG, medical imaging, pilot study, Image, optimization, low-dose, F-FDG
News Source: Blake Davidson. (November 10, 2026). Deep Learning Enhances Low-Dose Breast PET/MRI Image Quality. Scienmag.
