Deep Denoising of O-15 Water Dynamic PET Images without Training Data

Objectives: O-15 labeled water is a fast-decaying PET tracer for quantitative blood flow measurement. Due to its short half-life and low accumulation, O-15 water dynamic PET images are often very noisy and suboptimal for visual assessment. Averaging of repeated injections could be used for noise reduction, but the performance is limited. Despite of the superior performance of deep learning-based denoising on medical images, there is hardly any high-quality O-15 water images to provide clean training labels. This study aimed to generate high-quality O-15 water dynamic images using deep neural networks without additional training data.

Methods: Our method utilized the Noise2Noise training schemes for deep neural networks, in which the network can be trained in the denoising process. And there is no additional training input required besides two noisy images. Because different injections in the O-15 water scans have independent noises, we trained an encoder-decoder structured 3D convolutional neural network to map time frames across injections. After the network was trained, it served as a frame-by-frame denoising network for the dynamic PET images. Three losses were used to train the network: (1) Denoising: L2 loss between output and noisy frames from other injections; (2) Bias control: low pass-filtered L2 loss between output and input to compensate for the bias due to variability across injections; (3) Content: L2 loss between output and input to improve edges. The network was optimized for the weighted average of the three losses. Because of the relatively high dimension of dynamic PET, no overfitting to noise was observed even when only one dataset was used. We validated the method with an O-15 water study of a rabbit who had a VX-2 tumor xenograft in its flank. The scan was performed on Siemens Biograph with 3 injections of O-15 water (~5 mCi per injection). Each scan lasted for 198 seconds with 6 seconds per frame. The dynamic images were denoised frame by frame and one-tissue compartment model was used to calculate pixelwise K₁, which is closely related to blood flow. We also implemented time-intensity profile similarity (TIPS), which is a spatial-temporal denoising algorithm for dynamic images. The information from all 3 injections was used in TIPS for fair comparison.

Results: Noise2Noise significantly reduced the noise of dynamic images (Fig. 1). Compared to TIPS, Noise2Noise prevented over-smoothing in colder areas such as extremities. It also preserved details such as a metastasis from the primary tumor, which was almost smoothed out in the TIPS results. On K₁, Noise2Noise had reduced noise compared to original images and more details compared to TIPS. We noticed that both Noise2Noise and TIPS had considerably lower K₁ values compared to original images. It was due to the noise-induced bias when performing regression. Contrast recovery (CR) of the primary tumor and standard deviation in the liver were calculated for the K₁ images with various denoising hyperparameters (Fig. 2). Noise2Noise demonstrated higher tumor CR with lower liver standard deviation. The bias of the denoising methods were evaluated in each time frame, by comparing the mean tumor uptakes to that in the original images. Noise2Noise and TIPS had relative biases of 4.10% ± 4.49% and 10.36% ± 3.87% respectively. We confirmed that the proposed method outperforms the conventional method qualitatively and quantitatively.

Conclusions: The proposed Noise2Noise denoising method could significantly reduce the noise in O-15 water dynamic PET images with structural preservation and low bias. It also led to better K₁ images. No additional training data were required and the better denoising performance was achieved compared to averaging images from multiple injections. The proposed method greatly improved the value for visual assessment of O-15 water PET scans.

• Download figure
• Open in new tab
• Download powerpoint