Source-linked AI summary
200x Low-dose PET Reconstruction using Deep Learning
Junshen Xu, Enhao Gong, John Pauly, Greg Zaharchuk
TL;DR
Ultra-low-dose PET reconstruction is difficult because reducing tracer dose lowers SNR and information while challenging conventional reconstruction. The paper uses a fully convolutional encoder-decoder residual network with concatenate skip connections and multi-slice inputs, achieving comparable standard-dose image quality from 0.5% of the regular dose.
Problem
Dose reduction lowers PET SNR and image quality, while prior methods had not reconstructed PET images at higher dose-reduction factors.
Method
A fully convolutional encoder-decoder residual network uses concatenate skip connections and multi-slice inputs for ultra-low-dose PET reconstruction.
Results
The method achieves significantly better reconstruction than previous methods using ultra-low-dose PET data from 0.5% of the regular dose.
Takeaways & Limitations
The reconstructed images have comparable quality to normal-dose PET, with reduced noise and preserved resolution and detailed structures.
Abstract
from arXiv · showhide
Positron emission tomography (PET) is widely used in various clinical applications, including cancer diagnosis, heart disease and neuro disorders. The use of radioactive tracer in PET imaging raises concerns due to the risk of radiation exposure. To minimize this potential risk in PET imaging, efforts have been made to reduce the amount of radio-tracer usage. However, lowing dose results in low Signal-to-Noise-Ratio (SNR) and loss of information, both of which will heavily affect clinical diagnosis. Besides, the ill-conditioning of low-dose PET image reconstruction makes it a difficult problem for iterative reconstruction algorithms. Previous methods proposed are typically complicated and slow, yet still cannot yield satisfactory results at significantly low dose. Here, we propose a deep learning method to resolve this issue with an encoder-decoder residual deep network with concatenate skip connections. Experiments shows the proposed method can reconstruct low-dose PET image to a standard-dose quality with only two-hundredth dose. Different cost functions for training model are explored. Multi-slice input strategy is introduced to provide the network with more structural information and make it more robust to noise. Evaluation on ultra-low-dose clinical data shows that the proposed method can achieve better result than the state-of-the-art methods and reconstruct images with comparable quality using only 0.5% of the original regular dose.
I. INTRODUCTION
PET dose reduction lowers image quality and makes reconstruction difficult, motivating a deep learning approach that reconstructs standard-dose images from ultra-low-dose inputs at DRF=200.
- Motivation: Standard-dose PET improves clinical imaging but exposes patients to radioactive-tracer radiation, motivating efforts to reduce administered dose.PET supports applications including cancer diagnosis, tumor detection, and early diagnosis of neurological diseases.
- Motivation: Dose reduction lowers PET image SNR and degrades image quality, which can impair diagnostic imaging.
- Prior work: Existing approaches include iterative reconstruction, image filtering, post-processing, and machine learning methods for improving low-dose PET images.Iterative methods combine acquired-data statistics with regularization, while machine learning methods can map paired low-dose images to standard-dose images.
- Research gap: Prior methods addressed standard-dose denoising or quarter-dose reconstruction, but no work had reconstructed low-dose PET images at a higher dose-reduction factor.
A. Dataset and experiments setup
The experiments use clinical PET/MRI data and synthesize DRF=200 inputs by randomly retaining 0.5% of recorded count events, then train the network to recover standard-dose reconstructions.
- Dataset: DRF=200 low-dose raw data were synthesized by randomly selecting 0.5% of count events uniformly across the acquisition period.
- Reconstruction: Standard-dose and DRF=200 PET images were reconstructed from the corresponding raw data using standard OSEM methods.
- Training objective: The training objective was to learn reconstruction from the DRF=200 image to the DRF=1 standard-dose reconstruction.
- Network architecture: The proposed network uses an encoder-decoder structure with symmetry concatenate connections between corresponding stages, inspired by U-Net but adapted for image synthesis.Stages use convolutions, batch normalization, and ReLU, with max pooling for downsampling and bilinear interpolation for upsampling.
2) Residual learning:
The method uses residual learning and multi-slice inputs to improve ultra-low-dose PET reconstruction, while selecting L1 loss for perceptual quality and efficient training.
- Residual learning: The network learns the difference between standard-dose outputs and low-dose inputs through a residual connection from input to output.Residual learning is reported to improve network performance for low-dose PET reconstruction.
- Multi-slice input: Three adjoining slices provide 2.5D structural information that helps distinguish random noise from consistent brain structure.The paper contrasts this with single-slice input, where noise may be hallucinated as structure.
- Loss-function selection: L1 loss is selected because it avoids patchy artifacts associated with L2 loss while adding little back-propagation overhead compared with SSIM and MS-SSIM.The paper also considers SSIM and MS-SSIM as structural and perceptual similarity measures.
C. Computation environment and hardware settings
The network was implemented in TensorFlow and trained on a two-GPU Ubuntu server using RMSprop for 120 epochs.
- Computation environment and hardware settings: Training used 2 NVIDIA GTX 1080Ti GPUs on an Ubuntu server with a TensorFlow implementation.The experiments used RMSprop with a learning rate decreasing from 1 × 10^-3 to 2.5 × 10^-4.
- Computation environment and hardware settings: The network was trained for 120 epochs with truncated Gaussian kernel initialization and zero-initialized biases.The kernels used zero mean and standard deviation 0.02.
D. Evaluation and similarity metrics
Evaluation uses leave-one-out cross-validation to assess generalization to new patients and three image-similarity metrics computed within a brain mask.
- Evaluation protocol: Leave-one-out cross-validation trains each patient reconstruction model on the other eight patients to quantify generalization error.The procedure is intended to evaluate performance on new patient data with different pathology.
- Similarity metrics: The study evaluates image quality with NRMSE, PSNR, and SSIM.SSIM is defined in the paper, while NRMSE and PSNR are also used for quantitative comparison.
- Similarity metrics: All similarity metrics are computed after applying a brain mask estimated from image support.This is intended to better match metric computation to clinical assessment.
III. RESULTS
The proposed method was evaluated against NLM, BM3D, and AC-Net using cross-validation, averaged metrics, representative slices, and visual comparisons.
- Comparison with other methods: Cross-validation compared the proposed method with NLM, BM3D, and AC-Net for low-dose PET reconstruction.The evaluation used LOOCV and included NRMSE, PSNR, and SSIM.
- Quantitative evaluation: Fig. 4 reports averaged NRMSE, PSNR, and SSIM across all subjects, while Fig. 5 reports these metrics for all nine leave-one-out tests.
- Visual evaluation: Representative-slice evaluation combined quantitative metrics with reconstruction images, zoomed tumors, and corresponding error maps.
B. Contributions of skip connections
The study examines residual and concatenate skip connections alongside multi-slice inputs, using ablations and visual comparisons to assess reconstruction behavior.
- Contributions of skip connections: Four models were compared: both skip-connection types, concatenate-only, residual-only, and no skip connections.Testing loss and cross-validation results were used to evaluate the alternatives.
- Contributions of multi-slice inputs: Multi-slice input combines adjoining slices as feature augmentation to reduce noise and artifacts while preserving structure and details.
- Contributions of skip connections: The visual comparisons include standard-dose, low-dose, NLM, BM3D, AC-Net, and proposed reconstructions, plus zoomed regions and error maps.
- Contributions of multi-slice inputs: Three-slice input performed significantly better than single-slice input, while adding more than three slices produced less significant gains.
D. Depth of network
The network depth was optimized by grid search, while broader comparisons assessed reconstruction quality and computational efficiency against other methods.
- Depth of network: A grid search over pooling layers and convolutions found np = 3 and nc = 2 as the best architecture in this study.The search varied np from 2 to 5 and nc from 1 to 3.
- Results: The proposed method achieved the best quantitative performance across all nine subjects and the highest reported perceptual quality in the evaluated comparisons.The GBM region also received the best perceptual result in the cited visual comparison.
- Computational costs: The proposed solution was reported as more accurate and more efficient than the compared methods.Inference benefits from TensorFlow implementation and GPU parallelization, although training requires substantial time.
- Computational costs: The encoder-decoder structure supports more parameters and channels for higher-level feature extraction while reducing computation time relative to the single-scale AC-Net model.
B. Benefits from concatenation and residual skip connections
Both concatenate and residual skip connections contribute to reconstruction quality, while three-slice input improves structural detail preservation without further gains from additional slices.
- B. Benefits from concatenation and residual skip connections: Both skip-connection types together achieve the best performance, whereas either type alone performs close to or worse than no skip connection.The result indicates that concatenate and residual connections are not independent.
- C. Benefits from 2.5D augmentation: Three-slice input preserves detailed structures that become blurred with single-slice input.Stacking adjoining slices provides additional structural information for reconstruction.
- C. Benefits from 2.5D augmentation: Adding more than three input slices does not further improve performance.The reported improvement from 2.5D augmentation appears with three slices, but not with larger multi-slice inputs.
V. CONCLUSION
The proposed fully convolutional network combines multiscale encoding and decoding with concatenate and residual connections for ultra-low-dose PET reconstruction. It produces comparable normal-dose quality, preserves detail while reducing noise, and outperforms previous methods using 0.5% of regular dose.
- V. CONCLUSION: The proposed method uses a multiscale encoder-decoder architecture with concatenate connections and residual learning.These components form the deep fully convolutional network for ultra-low-dose PET reconstruction.
- V. CONCLUSION: The method generates PET images with quality comparable to normal-dose images while reducing noise and preserving resolution and detailed structures.The reported evaluation describes both denoising and detail preservation.
- V. CONCLUSION: The loss function, 2.5D multi-slice inputs, and concatenate and residual skip connections each contribute to improved performance.The paper evaluates these components through quantitative and qualitative comparisons.
- V. CONCLUSION: The method better preserves structure and avoids hallucination from noise and artifacts than the compared alternatives.This conclusion is based on detailed quantitative and qualitative comparisons.
- V. CONCLUSION: 0.5% of regular dose yields significantly better reconstruction than previous methods in the reported ultra-low-dose PET comparison.The paper frames this dose level as potentially enabling safer and more efficient PET scans.