4364
Low Dose PET/MR Imaging in Crohn's Disease1Department of Radiology, University of Wisconsin-Madison, Madison, WI, United States, 2Department of Electrical and Computer Engineering, University of Wisconsin-Madison, Madison, WI, United States, 3Department of Medical Physics, University of Wisconsin-Madison, Madison, WI, United States
Synopsis
For patients with Crohn’s Disease, reduction of radiation exposure during imaging exams is critical as life-long surveillance is often required. This research aims to reconstruct full dose PET images from low dose and MR images using a deep learning approach. Based on the analysis in twenty patients utilizing Leave-One-Out-Cross-Validation, a 10x reduction from clinical FDG tracer dose is attainable.
Introduction
Crohn's Disease (CD), an inflammatory bowel disease, will cause pain, diarrhea, fever, and even bleeding. Monitoring the activity of CD will facilitate diagnosis, measuring treatment response and tracking disease progression. Due to high sensitivity and specificity, simultaneous positron emission tomography-magnetic resonance imaging (PET/MRI) can help visualize fine anatomic details and reveal co-localized active inflammatory lesions. However, ionizing radiation exposure from fluorodeoxyglucose (FDG), the radioactive tracer, is a concern, since CD patients are young and will likely require life-long surveillance. Hence, the development of new low dose PET imaging techniques will be beneficial for monitoring CD and reduce the unnecessary risk for patients.
Methods
The twenty CD patients were imaged on a Signa PET/MR scanner (GE Healthcare, Waukesha, WI) with 5 mCi/70 kg of FDG (half of the standard clinical FDG dose). Two-point Dixon MR images, including water, fat, in-phase and out-of-phase were acquired and used for MR-based attenuation correction for five-minute PET images. Based on five-minute full dose images, one-minute low dose images were reconstructed to simulate a 1 mCi/70 kg injection 1.
A U-Net model 2 was utilized to reconstruct full dose data from low dose and MR data. Three variation inputs included three adjacent PET slices (PET-only), one MR slice (MR-only), or a concatenation of them (PET-MR). Inputs moved through a 5-level U-Net with 64 starting filters for 400 training epochs. Adam optimizer and Dropout were applied, as well as data augmentation [4] (2D-flip, zoom, 2D-shift, shear, and rotation). To evaluate the model, Leave-One-Out-Cross-Validation (LOOCV) was utilized. Each subject was tested by a model trained from 14 subjects and validated from 5 subjects. Three quantitative metrics, containing Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Normalized Root-Mean-Square-Error (NRMSE) were assessed.
Results & Discussion
As shown in Figure 1 and Table 1, among three metrics, PET-MR and PET-only models demonstrated better image qualities than the original Low Dose data. The MR-only model did not perform well in comparison. Bold entries in Table 1 indicates measures were statistically different at p<0.05 using a paired T-test in the whole body (All) and bowels regions (Seg). In a brief conclusion, PET-only and PET-MR models performed similarly well in the reconstruction task, while the MR models did not improve the image quality. Figures 2, 3, and 4 demonstrate the approach in three subjects. The bright signal in the bowels is a sign of inflammation. PET-only and PET-MR models reserve the clinical information with enhancement in image quality, while MR-only model fails to demonstrate any inflammation.Conclusion
Compared to a full clinical dose of FDG (10 mCi/70 kg), deep learning reconstruction enabled a 10x reduction of FDG tracer with similar even higher image qualities, measured quantitatively in PSNR, SSIM, and NRMSE. While the models utilizing PET-only and PET-MR provided similar performance, the MR-only model offered poor performance. This suggests that no information present within the utilized MR images would directly inform inflammation as seen on PET. While further research will study the application of additional MR sequences in similar approaches, this outcome is suggestive of the complementary nature of the physiological information provided by simultaneous PET/MR.Acknowledgements
Department of Radiology, University of Wisconsin
NIH R01 EB026708
References
1. Oehmigen, Mark, et al. "Radiotracer dose reduction in integrated PET/MR: implications from national electrical manufacturers association phantom studies." J Nucl Med 55.8 (2014): 1361-7.
2. Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation."International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015.
Figures
Figure 1. LOOCV results for three metrics based on models of different inputs
