Table 3 List of identified image quality improvement studies along with their main characteristics
From: Deep learning-enhanced nuclear medicine SPECT imaging applied to cardiac studies
No. | First author | Year | Title | Input | Learning algorithm | Outcome | Validation (total num. of images) | Results |
|---|---|---|---|---|---|---|---|---|
Low-dose | ||||||||
1 | Ramon | 2018 | Initial investigation of low-dose SPECT-MPI via deep learning | SPECT: Low-dose images (1/8 dose) | 3D-CNN | SPECT: Full-dose images | (Total num images: 930) | FBP: 0.99 O-SEM: 0.99 |
2 | Olia | 2021 | Deep learning–based denoising of low‑dose SPECT myocardial perfusion images: quantitative assessment and clinical performance | SPECT: Low-dose images (½ dose) | Generative adversarial network | SPECT: Full-dose images | 35 patients as an external dataset (Total num images: 345) | PSNR: 42.49 ± 2.37 SSIM: 0.99 ± 0.01 RMSE: 1.99 ± 0.63 R: 0.997 ± 0.001 |
3 | Ramon | 2020 | Improving diagnostic accuracy in low-dose SPECT myocardial perfusion imaging with convolutional denoising networks | SPECT: Low-dose images (½ dose) | 3D CNN | SPECT: Full-dose images | 122 patients for validation (Total num images: 1052) | AUC: 0.799 |
4 | Song | 2019 | LOW-DOSE CARDIAC-GATED SPECT STUDIES USING A RESIDUAL CONVOLUTIONAL NEURAL NETWORK | SPECT: Low-dose images (¼ dose) | 3D residual convolutional neural network (CNN | SPECT: Full-dose images | 12 patients for validation (Total num images: 119) | NMSE = 0.153 |
5 | Song | 2020 | Low-dose cardiac-gated spect via a spatiotemporal convolutional neural network | SPECT: Low-dose images (¼ dose) | Spatiotemporal CNN | SPECT: Full-dose images | 12 cases for validation (Total num images: 119) | (NMSE = 0.273) |
Fast-Scan | ||||||||
6 | Shiri | 2020 | Standard SPECT myocardial perfusion estimation from half-time acquisitions using deep convolutional residual neural networks | SPECT: Fast scan images | ResNet | SPECT: Standard time scan | 10FCV (Total num images: 363) | RMSE: 6.8 ± 2.7 ARE: = 3.1 ± 1.1% SSIM: = 0.97 ± 1.1 PSNR: = 36.0 ± 1.4 |
Attenuation-correction | ||||||||
7 | Yang | 2021 | Direct attenuation correction using deep learning for cardiac SPECT: a feasibility study | SPECT: NAC | DCNN | SPECT: AC | 10FCV (Total num images: 100) | Mean ± SD: 0.49% 6 4.35% Average absolute segmental errors: 3.31% 6 2.87% |
8 | Shi | 2020 | Deep learning-based attenuation map generation for myocardial perfusion SPECT | Both photopeak window and scatter window SPECT images + gender, age, height, weight and body mass index (BMI) | Deep fully convolutional neural networks | Attenuation maps | 25 testing images (Total num images: 65) | The NMAE between the reconstructed SPECT images that were corrected using the synthetic and CT-based attenuation maps was 0.26% ± 0.15%, whereas the localized absolute percentage error was 1.33% ± 3.80% in the left ventricle (LV) myocardium and 1.07% ± 2.58% in the LV blood pool |
9 | Mostafapour | 2021 | Deep learning-based attenuation correction in the image domain for myocardial perfusion SPECT imaging | SPECT: NAC | ResNet-U-Net | Attenuation maps | 19 SPECT non-AC images as external validation and 5% oc training was dedicated as validation dataset (Total num images: 99) | ME ResNet: 6.99 ± 16.72 SSIM ResNet: 0.99 ± 0.04 ME U-Net: − 4.41 ± 11.8 SSIM U-Net: 0.98 ± 0.05 |
10 | Chen | 2021 | Direct and indirect strategies of deep-learning-based attenuation correction for general purpose and dedicated cardiac SPECT | Both photopeak and scatter windows | U-Net and DuRDN | μ-maps | 20 cases for validation (Total num images: 270) | NMSE: 1.20 ± 0.72% |
11 | Liu | 2021 | Post-reconstruction attenuation correction for SPECT myocardial perfusion imaging facilitated by deep learning-based attenuation map generation | SPECT: NAC | 3D GAN | Attenuation maps | 30 cases for validation (Total num images: 30) | Global nMAE: 2 1.3% ± 8.0% Localized absolute percentage errors: 2 3.8% ± 4.5% |
12 | Hagio | 2022 | “Virtual” attenuation correction: improving stress myocardial perfusion SPECT imaging using deep learning | SPECT: NAC maps | based on U-Net | DLAC attenuation-corrected polar maps | 1865 for validation (Total num images: 11,532) | Test1: Sensitivity improved by 18.9% for DLAC accuracy improved by 13.1% Test2: specificity improved by 8% and accuracy by 5.5% |
13 | Chen | 2021 | CT-free attenuation correction for dedicated cardiac SPECT using a 3D dual squeeze-and-excitation residual dense network | Hybrid SPECT/CT stress/rest + BMI, gender, and scatter-window | 3D Dual Squeeze-and-Excitation Residual Dense Network (DuRDN) | SPECT AC | 40 for validation (Total num images: 172) | nRMSE: 2.01 ± 1.01%, |
14 | Torkaman | 2021 | Direct image-based attenuation correction using conditional generative adversarial network for SPECT myocardial perfusion imaging | SPECT: NAC | 3D cGAN | Attenuation maps | 5FCV (Total num images: 100) | nRMSE: 0.1410 ± 0.0768 |
15 | Chen | 2022 | Cross-vender, cross-tracer, and cross-protocol deep transfer learning for attenuation map generation of cardiac SPECT | SPECT: NAC | DuRDN | SPECT AC | 30 studies for validation (Total num images: 120) | nME: 1.11 ± 1.57% |
16 | Mostafapour | 2022 | Deep learning-guided attenuation correction in the image domain for myocardial perfusion SPECT imaging | SPECT: NAC | ResNet-U-Net | SPECT AC | 19 patients for external validation (Total num images: 99) | ResNet and U-Net models resulted in an ME of − 6.99 ± 16.72 and − 4.41 ± 11.8 and an SSI of 0.99 ± 0.04 and 0.98 ± 0.05 |
17 | Nguyen | 3D U-Net generative adversarial network for attenuation correction of SPECT images | SPECT: NAC | 3DU-Net-Gan | SPECT AC | 279 studies for validation (Total num images: 603) | SSIM: 0.945% nME: 0.034 | |
18 | Shanbhag | 2022 | Deep learning-based attenuation correction improves diagnostic accuracy of cardiac SPECT | Short-axis NC and AC images | cGAN | SPECT: Full-dose images | Total num images: 5490 | AUC: 0.79 Normalcy rate: 70.4% |
19 | Hagio | 2021 | “Virtual” attenuation correction: improving stress myocardial perfusion SPECT imaging using deep learning | NAC | U-Net | Attenuation–corrected (DLAC) perfusion polar maps | Total num images: 11,532 | AUC: 0.827 |
Denoising | ||||||||
20 | Liu | 2020 | Deep learning with noise-to-noise training for denoising in SPECT myocardial perfusion imaging | SPECT | CU-Net | Denoised images | 121 for validation (Total num images: 895) | PSNR was improved on average by 23% |
21 | Liu | 2021 | Improving detection accuracy of perfusion defect in standard dose SPECT-myocardial perfusion imaging by deep-learning denoising | SPECT | 3D CU-Net | Denoised images | 121 for validation (Total num images: 190) | AUC: 0.86 |
22 | Kikuchi | 2022 | A myocardial extraction method using deep learning for 99mTc myocardial perfusion SPECT images: A basic study to reduce the effects of extra-myocardial activity | SPECT | U-Net++ | Myocardial region extraction | 71 cases for validation (Total num images: 694) | Dice coefficient: 0.918 |
23 | Sun | 2022 | Dual gating myocardial perfusion SPECT denoising using a conditional generative adversarial network | SPECT | 3D cGAN | Denoised images | 6 cases for validation (Total num images: 20) | Best resolution with less blurring: 0.1671 ± 0.078 |
24 | Sun | 2019 | Generative adversarial network for denoising in dual gated myocardial perfusion SPECT using a population of phantoms and clinical data | SPECT/CT | GAN | Denoised images | (Total num images: 5 4D XCAT phantoms) | NSD:0.083 |
25 | Mok | 2018 | Initial investigation of using a generative adversarial network for denoising in dual gating myocardial perfusion SPECT | SPECT | GAN | Denoised images | (Total num images: 1152 frames) | NSD: 0.093 |
Reconstruction | ||||||||
26 | Song | 2019 | Approximate 4D reconstruction of cardiac-gated spect images using a residual convolutional neural network | SPECT | 3D residual CNN | SPECT reconstructed images | 20 cases for validation (Total num images: 197) | NMSE = 0.033 |
27 | Xie | 2022 | Increasing angular sampling through deep learning for stationary cardiac SPECT image reconstruction | 4D extended cardiac-torso phantoms heart images | U-Net | Generated synthetic four-angle images | (Total num images: 8 pig studies and two physical phantoms, and 20 human studies) | SD: 0.68 |
28 | Dietze | 2019 | Accelerated SPECT image reconstruction with FBP and an image enhancement convolutional neural network | SPECT/CT scans | CNN | SPECT reconstructed images | 20 cases for validation (Total num images: 128) | CNR: 12.5 |
29 | Chrysostomou | 2020 | SPECT imaging reconstruction method based on deep convolutional neural network | Sinograms 192 × 128 × 1 | CNN | Original ”true” activity the size of 128 × 128 × 1 | 60,000 for validation (Total num images: 600,000 phantoms) | Medium Noise: MSE: 0.003 MAE: 0.023 SSIM: 0.938 PCC: 0.971 |
Other imaging improvements | ||||||||
30 | Cheng | 2022 | Super-resolution reconstruction for parallel-beam SPECT based on deep learning and transfer learning: a preliminary simulation study | LR sinogram phantoms images | Residual network and transfer learning | High Resolution reconstructed SPECT images | three sets of XCAT phantoms (Total num images: 100,000 Shepp Logan-like phantoms, 100,000 Derenzo-like phantoms, and 100,000 Jaszczak-like phantoms) | SSIM: 0.9777 |
31 | Zhang | 2021 | A novel deep-learning-based approach for automatic reorientation of 3D cardiac SPECT images | SPECT/CT | STN (spatial transformer network) | Cardiac SPECT reorientation | 676 cases for validation (Total num images: 6762) | Center A R^2: 0.9926 Center B: R^2: 0.9823 |