[论文解读] 200x Low-dose PET Reconstruction using Deep Learning

Junshen Xu, Enhao Gong+2 more|arXiv (Cornell University)|Dec 12, 2017

Medical Imaging Techniques and Applications参考文献 8被引用 109

一句话总结

一个完全卷积的编码-解码器网络，具有残差学习和2.5D多切片输入，从超低剂量（DRF=200）数据重建标准剂量PET，性能优于最先进的去噪方法。

ABSTRACT

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.

研究动机与目标

• Motivate reducing radioactive tracer dose in PET without compromising diagnostic quality.
• Develop a deep learning model to reconstruct standard-dose PET from ultra-low-dose data (DRF=200).
• Leverage multi-slice (2.5D) inputs and residual/concatenate skip connections to preserve structure and reduce noise.
• Compare the proposed method against state-of-the-art denoising/reconstruction techniques on in-vivo data.
• Analyze the impact of loss functions, input slices, and network depth on reconstruction quality.

提出的方法

• Use a fully convolutional encoder-decoder with symmetric concatenate (U-Net–like) skip connections.
• Incorporate residual learning by adding a direct input-to-output skip connection to model the residual between low- and standard-dose images.
• Adopt multi-slice (2.5D) inputs by stacking adjacent slices as input channels to provide structural context.
• Train with L1 loss to improve perceptual quality and reduce patchy artifacts.
• Evaluate with leave-one-out cross-validation (LOOCV) on nine glioblastoma patient PET/MRI datasets.
• Compare against NLM, BM3D, and AC-Net denoising methods, and analyze the contribution of skip connections, multi-slice inputs, and network depth.

实验结果

研究问题

• RQ1 Can a DL model reconstruct standard-dose PET from ultra-low-dose (DRF=200) data with clinically acceptable quality?
• RQ2 What is the contribution of residual and concatenate skip connections to reconstruction performance?
• RQ3 Does 2.5D (multi-slice) input improve robustness to noise and preserve structure compared to single-slice input?
• RQ4 How does the network depth affect reconstruction quality and what is the optimal depth for this task?

主要发现

切片	NRMSE	PSNR	SSIM
切片 A - 低剂量	0.228	30.05	0.917
切片 A - NLM	0.153	33.50	0.958
切片 A - BM3D	0.145	32.98	0.961
切片 A - AC-Net	0.147	33.83	0.960
切片 A - 提议方法	0.124	35.33	0.974
切片 B - 低剂量	0.214	29.58	0.899
切片 B - NLM	0.142	33.13	0.947
切片 B - BM3D	0.189	30.70	0.926
切片 B - AC-Net	0.137	33.50	0.958
切片 B - 提议方法	0.106	35.66	0.969

• The proposed method achieves the best quantitative performance across all nine subjects versus NLM, BM3D, and AC-Net (LOOCV).
• Across representative slices, the proposed method yields lower NRMSE, higher PSNR, and higher SSIM than competing methods (Slice A: low-dose 0.228/30.05/0.917; NLM 0.153/33.50/0.958; BM3D 0.145/32.98/0.961; AC-Net 0.147/33.83/0.960; Proposed 0.124/35.33/0.974; Slice B: low-dose 0.214/29.58/0.899; NLM 0.142/33.13/0.947; BM3D 0.189/30.70/0.926; AC-Net 0.137/33.50/0.958; Proposed 0.106/35.66/0.969).
• The model with both residual and concatenate skip connections performs best; removing either type degrades performance.
• Using three-slice (2.5D) inputs significantly improves results over single-slice inputs; adding more than three slices yields diminishing returns.
• Depth analysis shows an architecture with three pooling layers and two convolutions between poolings (n_p=3, n_c=2) yields the best average NRMSE/PSNR/SSIM.
• The network demonstrates favorable inference speed (average 19 ms per 256x256 image) compared to traditional denoising methods.