Evaluation of a 2D UNet-Based Attenuation Correction Methodology for PET/MR Brain Studies

Utilizing deep learning techniques to improve image quality and noise reduction in preclinical low-dose PET images in the sinogram domain

BACKGROUND

Low-dose positron emission tomography (LD-PET) imaging is commonly employed in preclinical research to minimize radiation exposure to animal subjects. However, LD-PET images often exhibit poor quality and high noise levels due to the low signal-to-noise ratio. Deep learning (DL) techniques such as generative adversarial networks (GANs) and convolutional neural network (CNN) have the capability to enhance the quality of images derived from noisy or low-quality PET data, which encodes critical information about radioactivity distribution in the body.

PURPOSE

Our objective was to optimize the image quality and reduce noise in preclinical PET images by utilizing the sinogram domain as input for DL models, resulting in improved image quality as compared to LD-PET images.

METHODS

A GAN and CNN model were utilized to predict high-dose (HD) preclinical PET sinograms from the corresponding LD preclinical PET sinograms. In order to generate the datasets, experiments were conducted on micro-phantoms, animal subjects (rats), and virtual simulations. The quality of DL-generated images was weighted by performing the following quantitative measures: structural similarity index measure (SSIM), root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR). Additionally, DL input and output were both subjected to a spatial resolution calculation of full width half maximum (FWHM) and full width tenth maximum (FWTM). DL outcomes were then compared with the conventional denoising algorithms such as non-local means (NLM), block-matching, and 3D filtering (BM3D).

RESULTS

The DL models effectively learned image features and produced high-quality images, as reflected in the quantitative metrics. Notably, the FWHM and FWTM values of DL PET images exhibited significantly improved accuracy compared to LD, NLM, and BM3D PET images, and just as precise as HD PET images. The MSE loss underscored the excellent performance of the models, indicating that the models performed well. To further improve the training, the generator loss (G loss) was increased to a value higher than the discriminator loss (D loss), thereby achieving convergence in the GAN model.

CONCLUSIONS

The sinograms generated by the GAN network closely resembled real HD preclinical PET sinograms and were more realistic than LD. There was a noticeable improvement in image quality and noise factor in the predicted HD images. Importantly, DL networks did not fully compromise the spatial resolution of the images.

Direct attenuation correction for 99mTc-3PRGD2 chest SPECT lung cancer images using deep learning

Introduction

The attenuation correction technique of single photon emission computed tomography (SPECT) images is essential for early diagnosis, therapeutic evaluation, and pharmacokinetic studies of lung cancer. ^99mTc-3PRGD₂ is a novel radiotracer for the early diagnosis and evaluation of treatment effects of lung cancer. This study preliminary discusses the deep learning method to directly correct the attenuation of ^99mTc-3PRGD₂ chest SPECT images.

Methods

Retrospective analysis was performed on 53 patients with pathological diagnosis of lung cancer who received ^99mTc-3PRGD₂ chest SPECT/CT. All patients' SPECT/CT images were reconstructed with CT attenuation correction (CT-AC) and without attenuation correction (NAC). The CT-AC image was used as the reference standard (Ground Truth) to train the attenuation correction (DL-AC) SPECT image model using deep learning. A total of 48 of 53 cases were divided randomly into the training set, the remaining 5 were divided into the testing set. Using 3D Unet neural network, the mean square error loss function (MSELoss) of 0.0001 was selected. A testing set is used to evaluate the model quality, using the SPECT image quality evaluation and quantitative analysis of lung lesions tumor-to-background (T/B).

Results

SPECT imaging quality metrics between DL-AC and CT-AC including mean absolute error (MAE), mean-square error (MSE), peak signal-to-noise ratio (PSNR), structural similarity (SSIM), normalized root mean square error (NRMSE), and normalized Mutual Information (NMI) of the testing set are 2.62 ± 0.45, 58.5 ± 14.85, 45.67 ± 2.80, 0.82 ± 0.02, 0.07 ± 0.04, and 1.58 ± 0.06, respectively. These results indicate PSNR > 42, SSIM > 0.8, and NRMSE < 0.11. Lung lesions T/B (maximum) of CT-AC and DL-AC groups are 4.36 ± 3.52 and 4.33 ± 3.09, respectively (p = 0.81). There are no significant differences between two attenuation correction methods.

Conclusion

Our preliminary research results indicate that using the DL-AC method to directly correct ^99mTc-3PRGD₂ chest SPECT images is highly accurate and feasible for SPECT without configuration with CT or treatment effect evaluation using multiple SPECT/CT scans.

Classical and Learned MR to Pseudo-CT Mappings for Accurate Transcranial Ultrasound Simulation

Model-based treatment planning for transcranial ultrasound therapy typically involves mapping the acoustic properties of the skull from an X-ray computed tomography (CT) image of the head. Here, three methods for generating pseudo-CT (pCT) images from magnetic resonance (MR) images were compared as an alternative to CT. A convolutional neural network (U-Net) was trained on paired MR-CT images to generate pCT T images from either T1-weighted or zero-echo time (ZTE) MR images (denoted tCT and zCT, respectively). A direct mapping from ZTE to pCT was also implemented (denoted cCT). When comparing the pCT and ground-truth CT images for the test set, the mean absolute error was 133, 83, and 145 Hounsfield units (HU) across the whole head, and 398, 222, and 336 HU within the skull for the tCT, zCT, and cCT images, respectively. Ultrasound simulations were also performed using the generated pCT images and compared to simulations based on CT. An annular array transducer was used targeting the visual or motor cortex. The mean differences in the simulated focal pressure, focal position, and focal volume were 9.9%, 1.5 mm, and 15.1% for simulations based on the tCT images; 5.7%, 0.6 mm, and 5.7% for the zCT; and 6.7%, 0.9 mm, and 12.1% for the cCT. The improved results for images mapped from ZTE highlight the advantage of using imaging sequences, which improves the contrast of the skull bone. Overall, these results demonstrate that acoustic simulations based on MR images can give comparable accuracy to those based on CT.