Deep learning image reconstruction technique based on sinogram with 99m Tc-3PRGD2 chest SPECT

Deep learning image reconstruction technique based on sinogram with 99m Tc-3PRGD2 chest SPECT | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Deep learning image reconstruction technique based on sinogram with 99m Tc-3PRGD2 chest SPECT Tong Wang, Xiaona Jin, Haiqun Xing, Yaping Luo, Fang Li This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3997053/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Purpose This study is to evaluate the accuracy of a deep learning reconstruction method based on sinogram with 99m Tc-3PRGD2 chest SPECT. The aim is to shorten the local SPECT scanning time by 50% while preserving the quality of the images, allowing for faster completion of full-body SPECT scanning. Materials and Methods The images were selected from 33 patients diagnosed with lung cancer both clinically and pathologically. The full-projection and half-projection reconstruction techniques were used to create SPECT tomographic images. All the projection images were used as the " Ground Truth ", and half of the images were used to create full-projection SPECT images. A training dataset 28 for the building model and a test dataset 5 were used to evaluate the image quality by measuring the image error of the test dataset. Result The evaluation results of the image quality for the 99m Tc-3PRGD2 chest SPECT images using the deep learning reconstruction method based on sinogram were based on 5 test datasets. The following metrics were calculated: mean absolute error (MAE), mean-square error (MSE), Peak signal to noise ratio (PSNR), structural similarity (SSIM), normalized root mean square error (NRSM), and normalized Mutual Information (NMI). The average values of PSNR and SSIM were found to be 46.43 ± 5.05 and 0.92 ± 0.02, respectively. The mean values for MAE, MSE, NRSM, and NMI were 1.04 ± 0.52, 9.54 ± 7.24, 0.07 ± 0.03, and 1.59 ± 0.04, respectively. Conclusion A novel approach to SPECT imaging involves using deep learning and selecting only half of the projections to reconstruct SPECT images directly from a sinogram. This technique has been shown to yield tomographic images of comparable quality to those obtained from full projection images while reducing scanning time for 99m Tc-3PRGD2 chest SPECT by 50%. 99mTc-3PRGD2 SPECT Sinogram Reconstruction Deep Learning Figures Figure 1 Figure 2 Figure 3 Background Single photon emission computed tomography (SPECT) is a functional molecular imaging technology that is both non-invasive and quantitative. It has proven to be an invaluable tool for clinical diagnosis, accurate staging, and efficacy evaluation, as well as integrated diagnosis and treatment of various diseases affecting the nervous system, cardiovascular system, and tumors. Although SPECT imaging is useful, it has limitations in hardware design which causes the low image resolution and detection sensitivity. This affects the signal-to-noise ratio negatively, and the imaging process takes a long time. Therefore, the fewer SPECT whole-body scans are done compared to PET devices. Malignant tumors are still a leading cause of death, and their tendency to spread, or metastasize, is a defining feature. As a result, whole-body scanning is crucial in staging tumors. A new type of tracer, 99m Tc-3PRGD2, can bind specifically to the integrin αvβ3 receptor with high selectivity and strong affinity and is used for detecting tumors, imaging angiogenesis, and evaluating responses to treatment. Several multicenter studies have indicated that 99mTc-3PRGD2 imaging exhibits high sensitivity and specificity for detecting lung malignancies, making it a valuable tool for tumor staging [ 1 , 2 , 3 ] . However, due to limitations in SPECT imaging speed, 99m Tc-3PRGD2 is primarily used for local rather than whole-body scanning. Scientists and experts have been working to improve the speed of SPECT scanning. They have been focusing on both hardware and software aspects to make the process faster and more efficient. On the hardware front, a new generation SPECT system has been developed that offers high sensitivity and better spatial resolution [ 4 ] . However, these systems are often expensive and only designed for specific organs, such as cardiac or brain SPECT systems, which limits their clinical application. As for SPECT image reconstruction, the several algorithms have been developed, with statistical iterative reconstruction algorithms like Maximum Likelihood Expectation Maximization (ML-EM) and Ordered Subset Expectation Maximization (OSEM) proving the most effective. These algorithms can also incorporate physical models to compensate for r-ray scatter and attenuation in tissue during imaging, thereby improving the SPECT system's resolution. In practical clinical settings, it has been observed that SPECT system resolution increases with the number of iterations, although noise in the image also increases. During the iterative image reconstruction process, noise correlation is introduced, which can be challenging to quantify and may sometimes lead to incorrect detection. However, due to the loss of radioactive counting information in image acquisition, the resolution improvement is still not ideal, and the scanning speed improvement is also limited. Deep learning methods offer a fundamental shift in existing image reconstruction techniques, using mathematics and physics. Molecular imaging specialists are increasingly interested in deep learning-based reconstruction technology, which can improve image quality by interpolating sparse sampling data or denoising limited sampling or low-dose images in the image or projection domain. Deep learning methods, such as established deep neural network structures like UNet, generative adversarial networks (GANs), and encoding-decoding structures, among others, take in raw projection sinogram data or low-quality reconstructed images obtained via traditional methods to the neural network. These methods use supervised or semi-supervised learning to output high-quality reconstructed images, improving scanning speed and image quality. Recent advancements in deep learning have significantly improved medical image segmentation [ 5 , 6 ] , particularly in MRI and CT imaging [ 7 ] . Researchers have explored using deep learning image reconstruction techniques in SPECT, inspired by these achievements. Ryden et al. employed deep learning to enhance the number of SPECT projections and improve scanning speed [ 8 ] . Shao, Chrysostomou, and Pan et al. studied various neural network structures to enhance the quality of brain, bone, and model SPECT images. Shiri et al. used the ResNet neural network to compare 99m Tc-sestamibi myocardial perfusion SPECT images [ 9 , 10 , 11 ] , reducing scanning time and the number of projections. They discovered that reducing the number of projections can better improve myocardial perfusion scanning imaging speed [ 12 ] . For the first time, a 3D UNet deep learning network was utilized in this study to directly reconstruct SPECT images from raw sinogram data using half projection. This was done to improve the scanning speed of 99m Tc-3PRGD2 chest SPECT and enable 99m Tc-3PRGD2 whole-body SPECT imaging. The aim was to increase the speed of SPECT scanning and pave the way for conducting whole-body SPECT scanning with 99m Tc-3PRGD2. Materials and Methods 2.1 Study Design , Selection of Participants, and Data Collection : Clinical ethics approval was obtained from Beijing Union Medical College Hospital to conduct a retrospective analysis of 33 patients with confirmed lung cancer between 2012 and 2017. The patients were randomly selected and included 17 males and 16 females, with an average age of 59.94 ± 7.20 years old and a lung lesion size of 36.49 ± 17.12mm. Each patient underwent a 99m Tc-3PRGD2 chest SPECT/CT scan, with an intravenous injection of 99m Tc-3PRGD2 11.1MBq (0.3mCi) per kilogram of body weight. The total dose administered ranged from 532.8 to 876.9 (679.3 ± 94.00) MBq per patient. The imaging was performed using the Philips Precision SPECT/CT system (Philips, Netherlands), with low-dose chest CT scans used to correct for attenuation in SPECT images. SPECT scanning parameters included a low energy high-resolution collimator, energy window 140 keV ± 15%, 360 ° scanning, 64 projections, projection/30 seconds, zoom 1.3, matrix 128 ◊128, and pixels 4.664mm. CT parameters included 120kVp/30 mAs, 512◊512 matrix, pixels 0.68mm, and slice thickness 3.0mm. 2.2 Data Preprocessing : Once the scanning is finished, the SPECT projection data and CT reconstructed images are sent to the image processing workstation. The original projection data is then transferred to a personal computer, and 32 projections are extracted using full projection and one projection interval to create sinograms and reconstruct SPECT images without attenuation correction (NAC). The NeuTomPy toolbox, an open-source software, is utilized for this process( https://neutompy-toolbox.readthedocs.io ) [ 13 ] . The filtered back projection FBP image reconstruction method (reconstruction method FBP_CUDA, filtering function hamming) is used. The reconstructed SPECT images are used as the reference standard (Ground Truth), and 28 training dataset are utilized for training the model, while 5 testdataset are used for testing. 2.3. Model training and testing : For this study, we used the MONAI deep learning framework and incorporated a 3D UNet deep neural network [ 14 ] , which is depicted in Fig. 1 . The dataset underwent preprocessing, and the image size was uniformly adjusted to (64, 64, 64). Additionally, we increased the sample size of the data by randomly rotating it 90°. The 3D Unet network parameters used in this study are listed below : (1)The number of network input channels is 1, which means the input is a 3D grayscale image. (2)The number of network output channels is also 1, and the output image size is (64, 64, 64). (3)After each convolution operation, the number of output channels in the network encoder section is (16, 32, 64, 128, 256), with a step size of 2. (4)The number of residual units is 3. (5)The normalization method is batch normalization. In this study, the mean square error loss function (MSELoss) was chosen as the loss function, with a preset value of 0.0001 for MSELoss. 2.4 SPECT image quality evaluation and statistical analysis : This study utilized quantitative analysis methods to assess the training model's effectiveness. Evaluation indicators were calculated based on established methods, such as mean absolute error (MAE), mean square error (MSE), peak signal-to-noise ratio (PSNR), structural similarity (SSIM), normalized root mean square error (NRMSE), normalized mutual information (NMI) of images. Reference material [ 12 , 15 ] supply detailed explanations on how to calculate these indicators. All statistical analyses for this study were performed with R version 3.5.1 and Python version 3.5.6. Continuous data was represented as mean ± standard deviation. Result In this study, we present a diagram (Fig. 2 ) that illustrates the training model's performance using MSELoss as a performance indicator. We observe that MSELoss has good convergence, and it stabilizes after 300 Epochs. Additionally, we supply Fig. 3 , which displays the absolute error diagram between the test set data and Ground Truth. The colored icons in the diagram represent the error size, measured in counts. From Fig. 3 , we note that the error value is related to the radioactive distribution of chest SPECT images, with high errors in organ and lesion counts showing high uptake. We also supply Table 1 , which shows the image quality evaluation results of 5 test dataset. The results show that the MAE, MSE, PSNR, SSIM, NRSM, and NMI (mean ± SD) were 1.04 ± 0.52, 9.54 ± 7.24, 46.43 ± 5.05, 0.92 ± 0.02, 0.07 ± 0.03, and 1.59 ± 0.04, respectively, with PSN averaging 46.43 and SSIM Table 1 Image quality evaluation results of 5 test dataset no MAE MSE PSNR SSIM NRSME NMI 1 1.92868 22.4593 49.2019 0.897792 0.054654 1.59651 2 1.02646 6.49192 50.8287 0.913294 0.046156 1.6089 3 0.667841 5.98046 41.0455 0.933797 0.114781 1.53544 4 0.907758 5.63805 50.2502 0.932151 0.047793 1.63061 5 0.688333 7.15257 40.8189 0.936361 0.110631 1.55867 mean ± SD 1.04 ± 0.52 9.54 ± 7.24 46.43 ± 5.05 0.92 ± 0.02 0.07 ± 0.03 1.59 ± 0.04 reaching 0.92. Discussion The use of new tumor-specific SPECT tracers, like 99m Tc-3PRGD2, in clinical settings requires a comprehensive evaluation of tumors. Merely conducting local SPECT imaging may not supply accurate results for clinical staging and efficacy assessment. Therefore, improving SPECT imaging speed is essential for enhancing the diagnostic process of nuclear medicine molecular imaging, particularly for conducting whole-body scans. When it comes to improving the quality of SPECT images, using deep learning is more effective than reducing image noise [ 16 , 17 ] . Ryden et al. utilized convolutional neural networks to increase projection numbers and enhance SPECT tomographic images, leading to a significant improvement in signal-to-noise ratio. However, this method still requires iterative reconstruction of tomographic images from projection data. Shao employed the Unet neural network to produce SPECT images from 99m Tc-MDP whole-body SPECT/CT scan data, which involves two steps [ 9 ] . They transform the input data into fully connected feature extraction, which adds complexity to the neural network. In their study, Chrysostomou et al. utilized CNN neural networks to analyze model projection data [ 10 ] and extended their research to different radioactive concentrations. Our results in chest SPECT images showed a similar SSIM to the experimental outcomes. However, our MSE was higher than the findings reported by Chrysostomou because 99m Tc-3PRGD2 has a high uptake not only in tumors but also in the liver, leading to errors in the liver and spleen (refer to the figure). Meanwhile, Pan et al. employed Unet [ 11 ] to investigate bone SPECT scanning by combining CT and SPECT images for training. As motion can significantly affect abdominal CT scans, we focused our research on SPECT images. The SSIM value found in this study was much higher compared to the SSIM value of 0.76 in Pan's research results. In Shiri's study on myocardial perfusion, the average SSIM value was 0.96, but the highest average PSNR value was 36 [ 12 ] , suggesting that there was no significant improvement in the signal-to-noise ratio of SPECT images. In this study, the 3D Unet was utilized to generate SPECT images directly from Sinogram, while FBP images were considered as the "Ground Truth". Despite having a small training set data, the average MSE was below 10, and the test group showed satisfactory results. FBP images were selected as the "Ground Truth" in this study because they offer better image resolution compared to iterative reconstruction SPECT images. This study explores a method to obtain full projection data for chest SPECT image reconstruction using only half of the projection data. Through deep learning, SPECT images are directly generated from the original projection Sinogram, which can save 50% of the scanning time and double the scanning speed. The resulting PSNR is 46.43, SSIM is 0.92, and MAE is approximately 1, indicating the reliability of this method. The results outperform earlier studies that considered both SSIM and PSNR and those reported by Shiri and Pan. These differences may be attributed to the use of different neural network structures and radioactive tracers for different organs. This highlights the importance of considering both acquisition parameters and radioactive tracers in deep learning research for nuclear medicine molecular imaging, which differs from CT and MR research. It's worth noting that this study has some limitations. Currently, the research is only based on Philips SPECT data, and there is a need for additional methodological research on the projection data of GE and Siemens SPECT equipment. Additionally, the training group for this study had limited data, so expanding the dataset for future studies is necessary. Based on the preliminary findings, the utilization of deep learning techniques to produce SPECT images directly from the 50% raw data sinogram of SPECT projection can generate high-quality images. This innovative approach has the ability to double the scanning speed of SPECT and offer exciting possibilities for comprehensive whole-body SPECT imaging. Declarations Ethics approval and consent to participate This retrospective study of existing patient data and images was approved by the institutional review board of Peking Union Medical College Hospital. The requirement for informed consent was waived by the institutional review board of Peking Union Medical College Hospital, because the patient data and images were existed in this retrospective study. All methods were carried out in accordance with relevant guidelines and regulations. Consent for publication NA. Availability of data and materials The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. Competing interests The authors declare that they have no competing interests. Funding This study was supported by the Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences (grant number: CF 2021-I2M-1-2-2022564) , and the National High Level Hospital Clinical Research Funding (grant number: 2022-PUMCH-C-003) . The funders in the research provided financial support. Authors’ contributions TW and XJ wrote this manuscript. Data collection and data analysis were performed by HX, YL and FL. All authors have been involved in drafting and revising the manuscript and approved the final version to be published. All authors read and approved the final manuscript. Acknowledgements Not Applicable. References Sun H, Zhang GJ, Lu HW, Wang XM. Research progress of 99mTc-3PRGD2 SPECT/CT in staging and efficacy evaluation of lung tumors. Journal of Molecular Imaging, 2021,44: 868-872. Zhu Z, Miao W, Li Q, Dai H, Ma Q, Wang F, Yang A, Jia B, Jing X, Liu S, Shi J, Liu Z, Zhao Z, Wang F, Li F. 99mTc-3PRGD2 for integrin receptor imaging of lung cancer: a multicenter study. J Nucl Med. 2012;53:716-22. Thai AA, Solomon BJ, Sequist LV, Gainor JF, Heist RS. Lung cancer. Lancet. 2021;398(10299):535–54. Garcia EV. Deep learning, another important tool for improving acquisition efficiency in SPECT MPI Imaging. J Nucl Cardiol. 2021;28:2780-2783. Wang T, Xin HG, Li YG, Wang SC, Li F, Jing HL. Deep learning-based automated segmentation of eight brain anatomical regions using head CT images in PET/CT. BMC Med Imaging. 2022; 22:99. Xing H, Zhang X, Nie Y, Wang S, Wang T, Jing H, Li F. A deep learning-based post-processing method for automated pulmonary lobe and airway trees segmentation using chest CT images in PET/CT. Quant Imaging Med Surg 2022;1210:4747-4757. Parakh A, Cao J, Pierce TT, Blake MA, Savage CA, Kambadakone AR. Sinogram-based deep learning image reconstruction technique in abdominal CT: image quality considerations. Eur Radiol. 2021;3111:8342-8353. Ryden T, van Essen M, Marin I, Svensson J and Bernhardt P. Deep-Learning Generation of Synthetic Intermediate Projections Improves 177Lu SPECT Images Reconstructed with Sparsely Acquired Projections. J Nucl Med 2021; 62(4):528–535. Shao W, Rowe SP, Du Y. SPECTnet: a deep learning neural network for SPECT image reconstruction. Ann Transl Med. 2021;9:819. Chrysostomou C, Koutsantonis L, Lemesios C, Papanicolas CN. Deep Convolutional Neural Network for Low Projection SPECT Imaging Reconstruction. IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC). 2020; 1:4. Pan BY, Qi N, Meng QY, Wang JC,Peng SY, Qi CX, Gong NJ. Ultra high-speed SPECT bone imaging enabled by a deep learning enhancement method: a proof of concept. EJNMMI Phys. 2022;9:4343. Shiri I, Sabet KA, Arabi H, et al. Standard SPECT myocardial perfusion estimation from half-time acquisitions using deep convolutional residual neural networks. J Nucl Cardiol. 2020;28:2761–79. Micieli D, Minniti T, Gorini G, “NeuTomPy toolbox, a Python package for tomographic data processing and reconstruction”. SoftwareX. 2019;9:60-264. Hu Y, Lv D, Jian S, Lang L, Cui C, Liang M, Song L, Li S, Wu Z. Comparative study of the quantitative accuracy of oncological PET imaging based on deep learning methods. Quant Imaging Med Surg. 2023;136:3760-3775. Sara U, Akter M and Uddin MS Image Quality Assessment through FSIM, SSIM, MSE and PSNR—A Comparative Study. Journal of Computer and Communications. 2019;7, 8-18. Liu J, Yang Y, Wernick MN, Pretorius PH, King MA. Deep learning with noise-to-noise training for denoising in SPECT myocardial perfusion imaging. Med Phys (Lanc). 2021;481:156–68. Shao W, Rowe SP, Du Y. Artificial intelligence in single photon emission computed tomography (SPECT) imaging: a narrative review. Ann Transl Med. 2021;99:820. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-3997053","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":283710419,"identity":"248f1091-a52d-45e8-8cbd-32915b78f46b","order_by":0,"name":"Tong Wang","email":"","orcid":"","institution":"Peking Union Medical College Hospital","correspondingAuthor":false,"prefix":"","firstName":"Tong","middleName":"","lastName":"Wang","suffix":""},{"id":283710420,"identity":"a3d5c5b3-5430-48a8-bbcd-4027745b68ac","order_by":1,"name":"Xiaona Jin","email":"","orcid":"","institution":"Peking Union Medical College Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiaona","middleName":"","lastName":"Jin","suffix":""},{"id":283710421,"identity":"6d1224ad-168a-495f-a7dd-89e2d4891389","order_by":2,"name":"Haiqun Xing","email":"","orcid":"","institution":"Peking Union Medical College Hospital","correspondingAuthor":false,"prefix":"","firstName":"Haiqun","middleName":"","lastName":"Xing","suffix":""},{"id":283710422,"identity":"f25cb25c-e17f-4168-806d-66804cab4136","order_by":3,"name":"Yaping Luo","email":"","orcid":"","institution":"Peking Union Medical College Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yaping","middleName":"","lastName":"Luo","suffix":""},{"id":283710423,"identity":"792786a3-7a51-4684-99de-1c96be713aaf","order_by":4,"name":"Fang Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAqklEQVRIiWNgGAWjYBACPmbGB0DKBsRmI04LGzOzAZBKI0ULA1jLYVK0sDMzfi74dT7P4NrhZw8Yau4Q5TBm6Zl9t4sNbqeZGzAce0aMFv4D0rw9txM33M5hk2BsOEycLb95e86RpoVNmufHARK1WPM2JBdL3k4zk0g4RoQWfv7DzLd5/tjl8d1OfibxoYYILWDA2MaQAGYkEKkBCP6QongUjIJRMApGHAAAVDgzW8DSZNIAAAAASUVORK5CYII=","orcid":"","institution":"Peking Union Medical College Hospital","correspondingAuthor":true,"prefix":"","firstName":"Fang","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2024-02-28 15:20:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3997053/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3997053/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":53581660,"identity":"94e64be3-5ef6-4b68-89d1-2aa34cc9ef8a","added_by":"auto","created_at":"2024-03-27 17:36:58","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":464104,"visible":true,"origin":"","legend":"\u003cp\u003ePatient with \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 Chest SPECT Image Process flow.\u003c/p\u003e\n\u003cp\u003eUpper row: Generate a Sinogram using 64 full projections and reconstruct coronal images.\u003c/p\u003e\n\u003cp\u003eBottom row: Generate a Sinogram using 32 partial projections and reconstruct coronal images.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-3997053/v1/4ea9b7a92bde044af98e63e3.png"},{"id":53581661,"identity":"cb7ff6c7-4afb-4f2f-b80a-aafa48bc49a6","added_by":"auto","created_at":"2024-03-27 17:36:58","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":217307,"visible":true,"origin":"","legend":"\u003cp\u003eLoss and MSE curves of the training model.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-3997053/v1/24e45b44f811dacc6f598941.png"},{"id":53581662,"identity":"1b7d1814-ca27-4aff-ba26-f573e718a09b","added_by":"auto","created_at":"2024-03-27 17:36:58","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":359001,"visible":true,"origin":"","legend":"\u003cp\u003eError diagram between test set data and Ground Truth.\u003c/p\u003e\n\u003cp\u003eThe colored icons in the error chart represent the size of the error, measured in radioactive counts.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-3997053/v1/ce9ed09c4c55920580d1b503.png"},{"id":57438329,"identity":"970b2619-c40d-4996-8377-ac950ae03681","added_by":"auto","created_at":"2024-05-30 17:01:20","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1437637,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3997053/v1/8bafb714-6bdc-450e-a9af-e63658cfd838.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep learning image reconstruction technique based on sinogram with 99m Tc-3PRGD2 chest SPECT","fulltext":[{"header":"Background","content":"\u003cp\u003eSingle photon emission computed tomography (SPECT) is a functional molecular imaging technology that is both non-invasive and quantitative. It has proven to be an invaluable tool for clinical diagnosis, accurate staging, and efficacy evaluation, as well as integrated diagnosis and treatment of various diseases affecting the nervous system, cardiovascular system, and tumors. Although SPECT imaging is useful, it has limitations in hardware design which causes the low image resolution and detection sensitivity. This affects the signal-to-noise ratio negatively, and the imaging process takes a long time. Therefore, the fewer SPECT whole-body scans are done compared to PET devices. Malignant tumors are still a leading cause of death, and their tendency to spread, or metastasize, is a defining feature. As a result, whole-body scanning is crucial in staging tumors. A new type of tracer, \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2, can bind specifically to the integrin αvβ3 receptor with high selectivity and strong affinity and is used for detecting tumors, imaging angiogenesis, and evaluating responses to treatment. Several multicenter studies have indicated that 99mTc-3PRGD2 imaging exhibits high sensitivity and specificity for detecting lung malignancies, making it a valuable tool for tumor staging\u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/sup\u003e. However, due to limitations in SPECT imaging speed, \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 is primarily used for local rather than whole-body scanning. Scientists and experts have been working to improve the speed of SPECT scanning. They have been focusing on both hardware and software aspects to make the process faster and more efficient. On the hardware front, a new generation SPECT system has been developed that offers high sensitivity and better spatial resolution\u003csup\u003e[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/sup\u003e. However, these systems are often expensive and only designed for specific organs, such as cardiac or brain SPECT systems, which limits their clinical application. As for SPECT image reconstruction, the several algorithms have been developed, with statistical iterative reconstruction algorithms like Maximum Likelihood Expectation Maximization (ML-EM) and Ordered Subset Expectation Maximization (OSEM) proving the most effective. These algorithms can also incorporate physical models to compensate for r-ray scatter and attenuation in tissue during imaging, thereby improving the SPECT system's resolution. In practical clinical settings, it has been observed that SPECT system resolution increases with the number of iterations, although noise in the image also increases. During the iterative image reconstruction process, noise correlation is introduced, which can be challenging to quantify and may sometimes lead to incorrect detection. However, due to the loss of radioactive counting information in image acquisition, the resolution improvement is still not ideal, and the scanning speed improvement is also limited. Deep learning methods offer a fundamental shift in existing image reconstruction techniques, using mathematics and physics. Molecular imaging specialists are increasingly interested in deep learning-based reconstruction technology, which can improve image quality by interpolating sparse sampling data or denoising limited sampling or low-dose images in the image or projection domain. Deep learning methods, such as established deep neural network structures like UNet, generative adversarial networks (GANs), and encoding-decoding structures, among others, take in raw projection sinogram data or low-quality reconstructed images obtained via traditional methods to the neural network. These methods use supervised or semi-supervised learning to output high-quality reconstructed images, improving scanning speed and image quality. Recent advancements in deep learning have significantly improved medical image segmentation\u003csup\u003e[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e, particularly in MRI and CT imaging\u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e. Researchers have explored using deep learning image reconstruction techniques in SPECT, inspired by these achievements. Ryden et al. employed deep learning to enhance the number of SPECT projections and improve scanning speed\u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e. Shao, Chrysostomou, and Pan et al. studied various neural network structures to enhance the quality of brain, bone, and model SPECT images. Shiri et al. used the ResNet neural network to compare \u003csup\u003e99m\u003c/sup\u003eTc-sestamibi myocardial perfusion SPECT images\u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e, reducing scanning time and the number of projections. They discovered that reducing the number of projections can better improve myocardial perfusion scanning imaging speed\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e. For the first time, a 3D UNet deep learning network was utilized in this study to directly reconstruct SPECT images from raw sinogram data using half projection. This was done to improve the scanning speed of \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 chest SPECT and enable \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 whole-body SPECT imaging. The aim was to increase the speed of SPECT scanning and pave the way for conducting whole-body SPECT scanning with \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e\u003cspan\u003e\u003cstrong\u003e2.1 Study Design\u003c/strong\u003e\u003c/span\u003e\u003cspan\u003e\u003cstrong\u003e, Selection of Participants, and Data Collection\u003c/strong\u003e: Clinical ethics approval was obtained from Beijing Union Medical College Hospital to conduct a retrospective analysis of 33 patients with confirmed lung cancer between 2012 and 2017. The patients were randomly selected and included 17 males and 16 females, with an average age of 59.94\u0026thinsp;\u0026plusmn;\u0026thinsp;7.20 years old and a lung lesion size of 36.49\u0026thinsp;\u0026plusmn;\u0026thinsp;17.12mm. Each patient underwent a \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 chest SPECT/CT scan, with an intravenous injection of\u0026nbsp;\u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 11.1MBq (0.3mCi) per kilogram of body weight. The total dose administered ranged from 532.8 to 876.9 (679.3\u0026thinsp;\u0026plusmn;\u0026thinsp;94.00) MBq per patient. The imaging was performed using the Philips Precision SPECT/CT system (Philips, Netherlands), with low-dose chest CT scans used to correct for attenuation in SPECT images. SPECT scanning parameters included a low energy high-resolution collimator, energy window 140 keV\u0026thinsp;\u0026plusmn;\u0026thinsp;15%, 360 \u0026deg; scanning, 64 projections, projection/30 seconds, zoom 1.3, matrix 128 \u0026loz;128, and pixels 4.664mm. CT parameters included 120kVp/30 mAs, 512\u0026loz;512 matrix, pixels 0.68mm, and slice thickness 3.0mm.\u003cbr\u003e\u003c/span\u003e\u003cspan\u003e\u003cstrong\u003e2.2 Data Preprocessing\u003c/strong\u003e: Once the scanning is finished, the SPECT projection data and CT reconstructed images are sent to the image processing workstation. The original projection data is then transferred to a personal computer, and 32 projections are extracted using full projection and one projection interval to create sinograms and reconstruct SPECT images without attenuation correction (NAC). The NeuTomPy toolbox, an open-source software, is utilized for this process(\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://neutompy-toolbox.readthedocs.io\u003c/span\u003e\u003c/span\u003e)\u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e. The filtered back projection FBP image reconstruction method (reconstruction method FBP_CUDA, filtering function hamming) is used. The reconstructed SPECT images are used as the reference standard (Ground Truth), and 28 training dataset are utilized for training the model, while 5 testdataset are used for testing.\u003cbr\u003e\u003c/span\u003e\u003cspan\u003e\u003cstrong\u003e2.3. Model training and testing\u003c/strong\u003e:\u003cbr\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003eFor this study, we used the MONAI deep learning framework and incorporated a 3D UNet deep neural network\u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e, which is depicted in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. The dataset underwent preprocessing, and the image size was uniformly adjusted to (64, 64, 64). Additionally, we increased the sample size of the data by randomly rotating it 90\u0026deg;. The 3D Unet network parameters used in this study are listed below :\u003c/p\u003e\n\u003cp\u003e(1)The number of network input channels is 1, which means the input is a 3D grayscale image.\u003c/p\u003e\n\u003cp\u003e(2)The number of network output channels is also 1, and the output image size is (64, 64, 64).\u003c/p\u003e\n\u003cp\u003e(3)After each convolution operation, the number of output channels in the network encoder section is (16, 32, 64, 128, 256), with a step size of 2.\u003c/p\u003e\n\u003cp\u003e(4)The number of residual units is 3.\u003c/p\u003e\n\u003cp\u003e(5)The normalization method is batch normalization.\u003c/p\u003e\n\u003cp\u003eIn this study, the mean square error loss function (MSELoss) was chosen as the loss function, with a preset value of 0.0001 for MSELoss.\u003c/p\u003e\n\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.4 SPECT image quality evaluation and statistical analysis :\u003c/h2\u003e\n \u003cp\u003e\u003cspan\u003eThis study utilized quantitative analysis methods to assess the training model\u0026apos;s effectiveness. Evaluation indicators were calculated based on established methods, such as mean absolute error (MAE), mean square error (MSE), peak signal-to-noise ratio (PSNR), structural similarity (SSIM), normalized root mean square error (NRMSE), normalized mutual information (NMI) of images. Reference material\u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e supply detailed explanations on how to calculate these indicators. All statistical analyses for this study were performed with R version 3.5.1 and Python version 3.5.6. Continuous data was represented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation.\u003cbr\u003e\u003c/span\u003e\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Result","content":"\u003cp\u003eIn this study, we present a diagram (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) that illustrates the training model's performance using MSELoss as a performance indicator. We observe that MSELoss has good convergence, and it stabilizes after 300 Epochs. Additionally, we supply Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, which displays the absolute error diagram between the test set data and Ground Truth. The colored icons in the diagram represent the error size, measured in counts. From Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, we note that the error value is related to the radioactive distribution of chest SPECT images, with high errors in organ and lesion counts showing high uptake. We also supply Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, which shows the image quality evaluation results of 5 test dataset. The results show that the MAE, MSE, PSNR, SSIM, NRSM, and NMI (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD) were 1.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.52, 9.54\u0026thinsp;\u0026plusmn;\u0026thinsp;7.24, 46.43\u0026thinsp;\u0026plusmn;\u0026thinsp;5.05, 0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02, 0.07\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03, and 1.59\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04, respectively, with PSN averaging 46.43 and SSIM\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eImage quality evaluation results of 5 test dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eno\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePSNR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSSIM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNRSME\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNMI\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.92868\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e22.4593\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e49.2019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.897792\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.054654\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.59651\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.02646\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6.49192\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e50.8287\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.913294\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.046156\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.6089\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.667841\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.98046\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e41.0455\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.933797\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.114781\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.53544\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.907758\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.63805\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e50.2502\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.932151\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.047793\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.63061\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.688333\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e7.15257\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e40.8189\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.936361\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.110631\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.55867\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9.54\u0026thinsp;\u0026plusmn;\u0026thinsp;7.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e46.43\u0026thinsp;\u0026plusmn;\u0026thinsp;5.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.07\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.59\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003ereaching 0.92.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe use of new tumor-specific SPECT tracers, like \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2, in clinical settings requires a comprehensive evaluation of tumors. Merely conducting local SPECT imaging may not supply accurate results for clinical staging and efficacy assessment. Therefore, improving SPECT imaging speed is essential for enhancing the diagnostic process of nuclear medicine molecular imaging, particularly for conducting whole-body scans. When it comes to improving the quality of SPECT images, using deep learning is more effective than reducing image noise \u003csup\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e. Ryden et al. utilized convolutional neural networks to increase projection numbers and enhance SPECT tomographic images, leading to a significant improvement in signal-to-noise ratio. However, this method still requires iterative reconstruction of tomographic images from projection data. Shao employed the Unet neural network to produce SPECT images from \u003csup\u003e99m\u003c/sup\u003eTc-MDP whole-body SPECT/CT scan data, which involves two steps \u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e. They transform the input data into fully connected feature extraction, which adds complexity to the neural network. In their study, Chrysostomou et al. utilized CNN neural networks to analyze model projection data\u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/sup\u003e and extended their research to different radioactive concentrations. Our results in chest SPECT images showed a similar SSIM to the experimental outcomes. However, our MSE was higher than the findings reported by Chrysostomou because \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 has a high uptake not only in tumors but also in the liver, leading to errors in the liver and spleen (refer to the figure). Meanwhile, Pan et al. employed Unet \u003csup\u003e[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e to investigate bone SPECT scanning by combining CT and SPECT images for training. As motion can significantly affect abdominal CT scans, we focused our research on SPECT images. The SSIM value found in this study was much higher compared to the SSIM value of 0.76 in Pan's research results. In Shiri's study on myocardial perfusion, the average SSIM value was 0.96, but the highest average PSNR value was 36\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e, suggesting that there was no significant improvement in the signal-to-noise ratio of SPECT images. In this study, the 3D Unet was utilized to generate SPECT images directly from Sinogram, while FBP images were considered as the \"Ground Truth\". Despite having a small training set data, the average MSE was below 10, and the test group showed satisfactory results. FBP images were selected as the \"Ground Truth\" in this study because they offer better image resolution compared to iterative reconstruction SPECT images.\u003c/p\u003e \u003cp\u003eThis study explores a method to obtain full projection data for chest SPECT image reconstruction using only half of the projection data. Through deep learning, SPECT images are directly generated from the original projection Sinogram, which can save 50% of the scanning time and double the scanning speed. The resulting PSNR is 46.43, SSIM is 0.92, and MAE is approximately 1, indicating the reliability of this method. The results outperform earlier studies that considered both SSIM and PSNR and those reported by Shiri and Pan. These differences may be attributed to the use of different neural network structures and radioactive tracers for different organs. This highlights the importance of considering both acquisition parameters and radioactive tracers in deep learning research for nuclear medicine molecular imaging, which differs from CT and MR research.\u003c/p\u003e \u003cp\u003eIt's worth noting that this study has some limitations. Currently, the research is only based on Philips SPECT data, and there is a need for additional methodological research on the projection data of GE and Siemens SPECT equipment. Additionally, the training group for this study had limited data, so expanding the dataset for future studies is necessary.\u003c/p\u003e \u003cp\u003eBased on the preliminary findings, the utilization of deep learning techniques to produce SPECT images directly from the 50% raw data sinogram of SPECT projection can generate high-quality images. This innovative approach has the ability to double the scanning speed of SPECT and offer exciting possibilities for comprehensive whole-body SPECT imaging.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eThis retrospective study of existing patient data and images was approved by the institutional review board of Peking Union Medical College Hospital. The requirement for informed consent was waived by the institutional review board of Peking Union Medical College Hospital, because the patient data and images were existed in this retrospective study. All methods were carried out in accordance with relevant guidelines and regulations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNA.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by the Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences (grant number: CF 2021-I2M-1-2-2022564) , and the National High Level Hospital Clinical Research Funding (grant number: 2022-PUMCH-C-003) . The funders in the research provided financial support.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTW and XJ wrote this manuscript. Data collection and data analysis were performed by HX, YL and FL. All authors have been involved in drafting and revising the manuscript and approved the final version to be published. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot Applicable.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e"},{"header":"References ","content":"\u003col\u003e\n\u003cli\u003eSun H, Zhang GJ, Lu HW, Wang XM. Research progress of 99mTc-3PRGD2 SPECT/CT in staging and efficacy evaluation of lung tumors. Journal of Molecular Imaging, 2021,44: 868-872. \u003c/li\u003e\n\u003cli\u003eZhu Z, Miao W, Li Q, Dai H, Ma Q, Wang F, Yang A, Jia B, Jing X, Liu S, Shi J, Liu Z, Zhao Z, Wang F, Li F. 99mTc-3PRGD2 for integrin receptor imaging of lung cancer: a multicenter study. J Nucl Med. 2012;53:716-22. \u003c/li\u003e\n\u003cli\u003eThai AA, Solomon BJ, Sequist LV, Gainor JF, Heist RS. Lung cancer. Lancet. 2021;398(10299):535\u0026ndash;54. \u003c/li\u003e\n\u003cli\u003eGarcia EV. Deep learning, another important tool for improving acquisition efficiency in SPECT MPI Imaging. J Nucl Cardiol. 2021;28:2780-2783. \u003c/li\u003e\n\u003cli\u003eWang T, Xin HG, Li YG, Wang SC, Li F, Jing HL. Deep learning-based automated segmentation of eight brain anatomical regions using head CT images in PET/CT. BMC Med Imaging. 2022; 22:99. \u003c/li\u003e\n\u003cli\u003eXing H, Zhang X, Nie Y, Wang S, Wang T, Jing H, Li F. A deep learning-based post-processing method for automated pulmonary lobe and airway trees segmentation using chest CT images in PET/CT. Quant Imaging Med Surg 2022;1210:4747-4757.\u003c/li\u003e\n\u003cli\u003eParakh A, Cao J, Pierce TT, Blake MA, Savage CA, Kambadakone AR. Sinogram-based deep learning image reconstruction technique in abdominal CT: image quality considerations. Eur Radiol. 2021;3111:8342-8353. \u003c/li\u003e\n\u003cli\u003eRyden T, van Essen M, Marin I, Svensson J and Bernhardt P. Deep-Learning Generation of Synthetic Intermediate Projections Improves 177Lu SPECT Images Reconstructed with Sparsely Acquired Projections. J Nucl Med 2021; 62(4):528\u0026ndash;535. \u003c/li\u003e\n\u003cli\u003eShao W, Rowe SP, Du Y. SPECTnet: a deep learning neural network for SPECT image reconstruction. Ann Transl Med. 2021;9:819. \u003c/li\u003e\n\u003cli\u003eChrysostomou C, Koutsantonis L, Lemesios C, Papanicolas CN. Deep Convolutional Neural Network for Low Projection SPECT Imaging Reconstruction. IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC). 2020; 1:4.\u003c/li\u003e\n\u003cli\u003ePan BY, Qi N, Meng QY, Wang JC,Peng SY, Qi CX, Gong NJ. Ultra high-speed SPECT bone imaging enabled by a deep learning enhancement method: a proof of concept. EJNMMI Phys. 2022;9:4343. \u003c/li\u003e\n\u003cli\u003eShiri I, Sabet KA, Arabi H, et al. Standard SPECT myocardial perfusion estimation from half-time acquisitions using deep convolutional residual neural networks. J Nucl Cardiol. 2020;28:2761\u0026ndash;79.\u003c/li\u003e\n\u003cli\u003eMicieli D, Minniti T, Gorini G, \u0026ldquo;NeuTomPy toolbox, a Python package for tomographic data processing and reconstruction\u0026rdquo;. SoftwareX. 2019;9:60-264. \u003c/li\u003e\n\u003cli\u003eHu Y, Lv D, Jian S, Lang L, Cui C, Liang M, Song L, Li S, Wu Z. Comparative study of the quantitative accuracy of oncological PET imaging based on deep learning methods. Quant Imaging Med Surg. 2023;136:3760-3775.\u003c/li\u003e\n\u003cli\u003eSara U, Akter M and Uddin MS Image Quality Assessment through FSIM, SSIM, MSE and PSNR\u0026mdash;A Comparative Study. Journal of Computer and Communications. 2019;7, 8-18. \u003c/li\u003e\n\u003cli\u003eLiu J, Yang Y, Wernick MN, Pretorius PH, King MA. Deep learning with noise-to-noise training for denoising in SPECT myocardial perfusion imaging. Med Phys (Lanc). 2021;481:156\u0026ndash;68. \u003c/li\u003e\n\u003cli\u003eShao W, Rowe SP, Du Y. Artificial intelligence in single photon emission computed tomography (SPECT) imaging: a narrative review. Ann Transl Med. 2021;99:820. \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"99mTc-3PRGD2, SPECT, Sinogram, Reconstruction, Deep Learning","lastPublishedDoi":"10.21203/rs.3.rs-3997053/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3997053/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003ePurpose\u003c/h2\u003e \u003cp\u003eThis study is to evaluate the accuracy of a deep learning reconstruction method based on sinogram with \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 chest SPECT. The aim is to shorten the local SPECT scanning time by 50% while preserving the quality of the images, allowing for faster completion of full-body SPECT scanning.\u003c/p\u003e\u003ch2\u003eMaterials and Methods\u003c/h2\u003e \u003cp\u003eThe images were selected from 33 patients diagnosed with lung cancer both clinically and pathologically. The full-projection and half-projection reconstruction techniques were used to create SPECT tomographic images. All the projection images were used as the \" Ground Truth \", and half of the images were used to create full-projection SPECT images. A training dataset 28 for the building model and a test dataset 5 were used to evaluate the image quality by measuring the image error of the test dataset.\u003c/p\u003e\u003ch2\u003eResult\u003c/h2\u003e \u003cp\u003eThe evaluation results of the image quality for the \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 chest SPECT images using the deep learning reconstruction method based on sinogram were based on 5 test datasets. The following metrics were calculated: mean absolute error (MAE), mean-square error (MSE), Peak signal to noise ratio (PSNR), structural similarity (SSIM), normalized root mean square error (NRSM), and normalized Mutual Information (NMI). The average values of PSNR and SSIM were found to be 46.43\u0026thinsp;\u0026plusmn;\u0026thinsp;5.05 and 0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02, respectively. The mean values for MAE, MSE, NRSM, and NMI were 1.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.52, 9.54\u0026thinsp;\u0026plusmn;\u0026thinsp;7.24, 0.07\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03, and 1.59\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04, respectively.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eA novel approach to SPECT imaging involves using deep learning and selecting only half of the projections to reconstruct SPECT images directly from a sinogram. This technique has been shown to yield tomographic images of comparable quality to those obtained from full projection images while reducing scanning time for \u003csup\u003e99m\u003c/sup\u003eTc-3PRGD2 chest SPECT by 50%.\u003c/p\u003e","manuscriptTitle":"Deep learning image reconstruction technique based on sinogram with 99m Tc-3PRGD2 chest SPECT","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-03-27 17:36:53","doi":"10.21203/rs.3.rs-3997053/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"6f62dbbf-71f2-4511-8493-d246cb12b4f2","owner":[],"postedDate":"March 27th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-05-30T16:53:12+00:00","versionOfRecord":[],"versionCreatedAt":"2024-03-27 17:36:53","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3997053","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3997053","identity":"rs-3997053","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}