Generating Pseudo-Subtracted Image in Dual-Energy Contrast-Enhanced Spectral Mammography Using Transfer Learning

Generating Pseudo-Subtracted Image in Dual-Energy Contrast-Enhanced Spectral Mammography Using Transfer Learning | 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 Generating Pseudo-Subtracted Image in Dual-Energy Contrast-Enhanced Spectral Mammography Using Transfer Learning Asma Khorshidifar, Ghazal Mostaghel, Kaveh Dastvareh, Yashar Ahmadyar, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6147032/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 Background : Dual-energy contrast-enhanced spectral mammography (CESM) enhances breast cancer detection but increases radiation exposure, especially for high-risk patients like BRCA1 mutation carriers. Additionally, the dual-energy acquisition process can be time-consuming. This study uses deep learning to convert low-energy images into subtracted images, reducing radiation and contrast-related risks, while also addressing the time consumption challenge of the traditional CESM procedure. Methods : The study utilized the Categorized Digital Database for Low-energy and Subtracted Contrast-Enhanced Spectral Mammography Images (CDD-CESM), which contains 7600 image pairs after augmentation. The dataset was divided into 70% for training and 30% for testing. CycleGAN's performance was evaluated and compared against U-Net, Pix2Pix, and ResNet18. Key metrics for comparison included Structural Similarity Index and Peak Signal-to-Noise Ratio. The models were tested for their ability to generate high-quality subtracted images without the need for paired training data. Results : CycleGAN outperformed U-Net, Pix2Pix, and ResNet18 in generating pseudo-subtracted images. The SSIM score of 0.961, close to that of real subtracted images, indicates that CycleGAN successfully preserves structural details. Additionally, CycleGAN achieved this performance at a lower computational cost and without the need for paired data. Conclusions : CycleGAN effectively generates pseudo-subtracted images from low-energy mammography data, presenting a viable alternative to dual-energy imaging. This method has the potential to reduce the need for additional imaging, minimize radiation exposure, and simplify imaging procedures. The high SSIM score highlights CycleGAN's ability to maintain strong structural similarities in the generated images, making it a promising tool for detecting lesions in mammography. CESM Deep learning Breast cancer Figures Figure 1 Figure 2 Figure 3 Introduction Breast cancer is the most common cancer in women. In the early stages, non-metastatic cases are curable in 70–80% of patients; however, advanced breast cancer with metastases is currently incurable. The disease is molecularly diverse, with features like HER2 activation, hormone receptor positivity, and BRCA mutations influencing treatment. The treatment involves a combination of surgery, radiation therapy, systemic therapies such as endocrine therapy, chemotherapy, anti-HER2 treatment, and immunotherapy [ 1 ]. Dual-energy Contrast Enhanced Spectral Mammography (CESM) is a reliable tool for breast cancer diagnosis. The studies have shown that CESM sensitivity is comparable to Breast Magnetic Resonance Imaging (BMRI) in detecting breast cancer. It also has a significantly shorter exam time which has made it an accessible alternative to BMRI as well as an invaluable method in breast cancer detection and staging. Studies have found that CESM offers equal specificity and superior sensitivity in comparison to MRI, and excellent performance in diagnosing the pathological response of breast cancer to neoadjuvant chemotherapy (NAC). Additionally, CESM is more affordable and better tolerated by patients; hence it has become a promising option for evaluating tumors in the breasts [ 2 , 3 ]. CESM in an imaging technique to achieve both anatomical and functional information of the breasts, by capturing two images per view at different energy levels. The first scan is a low-energy image which shows breast structure exactly similar to standard 2D mammography. The second scan is a subtracted image that highlights areas of contrast uptake that may indicate malignant neovascularization [ 4 ]. However, it is important to note that having two scans per view in CESM increases the radiation dose to the patient and can be time-consuming, which may limit its practical application. Furthermore, the contrast agent used in CESM can potentially harm patients by causing allergic reactions or other adverse effects. Replacing two scans per view with one can effectively decrease the radiation dose, reduce the risk of allergic reactions to the contrast agent, and shorten the acquisition time. By leveraging deep neural networks, it is possible to virtually generate the subtracted images using only the low-energy scans [ 5 ]. Ouyang et al. proposed a deep convolutional model to generate virtual PET images for brain tumor detection to tackle the challenges of 18F-FDG PET such as cost, radiation exposure, and accessibility. In this retrospective study, convolutional neural networks were trained using multiple MRI sequences to produce synthesized PET images. The study shows the potential of deep learning to provide reliable, diagnostic-quality PET images from MRI, results in a cost-effective and non-invasive alternative for brain neoplasm assessment [ 6 ]. Cai presents a CycleGAN-based method for translating unpaired CT and MRI brain scans for a higher quality tumor diagnosis. The model trained on the Kaggle dataset, and the model demonstrated effective image translation with high quality. CycleGAN’s strengths lie in its ability to handle unpaired data with potential for broader applications in medical image analysis [ 7 ]. Moreover, several review articles have shown that image-to-image translation has been successfully applied across various modalities, such as PET, SPECT, MRI, CT, and mammography that facilitate effective cross-modality translation for improved diagnosis and analysis [ 8 – 12 ]. The studies have shown transfer learning is a fast and effective tool in different task of medical imaging [ 13 – 19 ]. In this work, we propose a transfer learning approach to generate subtracted mammography images from low-energy mammography scans. Material and Methods a. Dataset and preprocessing The Categorized Digital Database for Low-energy and Subtracted Contrast-Enhanced Spectral Mammography images (CDD-CESM), used in this study, is publicly available. It comprises a total of 1003 low-energy images and 1003 subtracted images (available at: https://www.cancerimagingarchive.net/collection/cdd-cesm/ ). In this dataset, patients received a low-osmolar iodinated contrast agent at a dose of 1.5 ml/kg. Two minutes after administration, the craniocaudal (CC) and mediolateral-oblique (MLO) views were obtained. The first scan is a low-energy exposure with peak kV values from 26 to 31 kV, and the other is a high-energy exposure from 45 to 49 kV. CC and MLO views were obtained and then recombined and subtracted to diminish the background parenchyma after the contrast agent had been injected. In the preprocessing part, images were standardized by resizing them to a uniform 256x256 dimension. Furthermore, augmentation techniques were applied, using different rotations, which increased the number of images to 7,600 images. The images were split into a training and blind-testing dataset in a 70%-30% ratio. b. Model Architecture The Cycle-Consistent Generative Adversarial Network (CycleGAN) [ 20 ] was employed for the image translation task. Even though image data would seem relatively simple to obtain, in some scenarios, the same image would need to be translated into two different target domains; hence, CycleGAN can be used for this as it does not rely on paired images for training. This is particularly beneficial when paired datasets are not possible, which is frequently the case in medical imaging. The main design of CycleGAN is bidirectionality, such that two classes can be transformed, in this instance, from low-energy mammography images to subtracted mammography images, and the opposite way. The model will comprise two generators and two discriminators. The generators are trained to map images from one domain to another (low energy to subtracted and back), and the discriminators are trained to distinguish between real and generated images in each domain This method aids in making sure that the transformation looks similar on both domains (Fig. 1 ). There are two main losses that are used while training CycleGAN—adversarial loss and cycle consistency loss. The adversarial loss is used to promote generated images to be indistinguishable from real images, according to the discriminator. Cycle consistency loss is fundamental to guaranteeing that the transformation is bidirectional, which requires the transformation of an image from domain A (low-energy) to domain B (subtracted) back to domain A to resemble as closely as possible the original input image. This loss guarantees that the network does sensible transformations, i.e., the images keep the meaningful content across transformations. The architecture is shown in Fig. 2 . For this study, the CycleGAN model was trained using the Adam optimizer with a learning rate of 0.0002, batch size of 1, for 14 epochs. Also, the Pix2Pix, U-Net, and ResNet18 models were trained and tested under the same conditions. The study was conducted using Google Colab's NVIDIA A100 GPU with 40GB of memory, 6912 CUDA cores, and a memory bandwidth of 1.6 TB/s. This high-performance hardware enabled efficient deep learning model training and image processing which provides the computational power necessary to complete the study. c. Evaluation Strategy Evaluation of the model’s performance was conducted using the following metrics: 1. Mean Squared Error (MSE) Mean Squared Error (MSE) MSE measures the average squared difference between corresponding pixels of the original and generated images. $$\:MSE\:=\:\left(\frac{1}{N}\right)*\:sum\left({\left({I}_{i}-\:{I}_{i}^{{\prime\:}}\right)}^{2}\right)$$ 1 N is the total number of pixels. \(\:{I}_{i}\) is the pixel value of the original image at position i. \(\:{I}_{i}^{{\prime\:}}\) is the pixel value of the generated image at position i. 2. Peak Signal-to-Noise Ratio (PSNR) PSNR indicates the ratio between the maximum possible pixel value and the noise level. $$\:PSNR\:=\:20\:*\:log10(MAX\_I)\:-\:10\:*\:log10\left(MSE\right)$$ 2 MAX_I is the maximum possible pixel value of the image. 3. Structural Similarity Index (SSIM) SSIM measures the perceptual difference between the original and generated images, focusing on changes in structural information. SSIM values range between − 1 and 1, where 1 indicates perfect similarity. $$\:SSIM\:=\frac{\left[\left(2\:*\:{\mu\:}_{I}*\:{\mu\:}_{I}^{{\prime\:}}+\:C1\right)*\:\left(2\:*\:{\sigma\:}_{II}^{{\prime\:}}+\:C2\right)\right]}{\left[\left({\mu\:}_{I}^{2}+\:{\mu\:}_{I}^{{\prime\:}2}+\:C1\right)*\:\left({\sigma\:}_{I}^{2}+\:{\sigma\:}_{I}^{{\prime\:}2}+\:C2\right)\right]}$$ 3 \(\:{{\mu\:}}_{\text{I}}\text{a}\text{n}\text{d}\:{{\mu\:}}_{\text{I}}^{{\prime\:}}\) are the average pixel values of the original and generated images. \(\:{{\sigma\:}}_{\text{I}}^{2}\text{a}\text{n}\text{d}\:{{\sigma\:}}_{\text{I}}^{{\prime\:}2}\) are the variances of the original and generated images. \(\:\sigma\:\_II{\prime\:}\) is the covariance between the original and generated images. C1 and C2 are small constants to stabilize the division. 4. Mean Absolute Error (MAE) MAE measures the average absolute difference between corresponding pixels of the original and generated images $$\:MAE\:=\:\left(\frac{1}{N}\right)*\:sum\left(\left|{I}_{i}-\:{I}_{i}^{{\prime\:}}\right|\right)$$ 4 N is the total number of pixels. \(\:{I}_{i}\) is the pixel value of the original image at position i. \(\:{I}_{i}^{{\prime\:}}\) is the pixel value of the generated image at position i. Results The CycleGAN-generated subtracted images exhibited robust structural and visual similarity to CESM-derived images. These images successfully excluded extraneous features from low-energy images and increased contrast, which helped with the recognition of malignant lesions. Figures 3 contains sample images including the predicted subtracted images, the low-energy images, and the CESM-generated images The metrics for the 1140 unseen test samples are presented in Table 1 . Table 1 Evaluation metrics for different models. Model PSNR MAE MSE SSIM CycleGAN 17.962 \(\:\pm\:\) 2.158 0.085 \(\:\pm\:\) 0.003 0.007 \(\:\pm\:\) 0.0003 0.961 \(\:\pm\:\) 0.012 U-net 11.962 \(\:\pm\:\) 0.898 0.285 \(\:\pm\:\) 0.001 0.081 \(\:\pm\:\) 0.0010 0.793 \(\:\pm\:\) 0.006 Pix2Pix 15.962 \(\:\pm\:\) 0.235 0.135 \(\:\pm\:\) 0.001 0.018 \(\:\pm\:\) 0.0002 0.833 \(\:\pm\:\) 0.011 ResNet 18 6.367 \(\:\pm\:\) 0.015 0.452 \(\:\pm\:\) 0.026 0.204 \(\:\pm\:\) 0.0041 0.683 \(\:\pm\:\) 0.042 Discussion The results of this study show that CycleGAN may have utility as a possible technique for the generation of virtual subtracted images similar to CESM. In particular, CycleGAN assists with the detection of cancer lesions by diminishing clutter from low-energy input images and enhancing contrast. This means that dual-energy images are not required to generate high-quality subtracted images because CycleGAN is able to provide the end subtracted state given just a low-energy image as input. This enables imaging with minimal optimization and virtually no data collection infrastructure required to create dual-energy data, resulting in cost and practical advantages. This approach minimizes both the hardware requirements and operational costs when obtaining dual-energy images while decreasing the patient’s radiation exposure—an important consideration in medical imaging. CycleGAN is evaluated on 1140 unseen samples across multiple metrics. CycleGAN outperformed U-net, Pix2Pix, and ResNet18 across all evaluation metrics, notably for structural similarity and peak signal-to-noise ratio, which are key qualities in ensuring the generated images are similar in visual appearance and structure. CycleGAN has a 0.961 structural similarity index score, which shows that its structural similarity to ground truth images is relatively high and that its key features necessary for accurate lesion detection are preserved. PSNR is 17.962 ± 2.158, which further demonstrates that CycleGAN is highly capable of image generation. U-net, Pix2Pix, and ResNet18 produced markedly lower peak signal-to-noise ratio and structural similarity index scores, indicating that the models perform poorly in retaining the structural nuances necessary for the appropriate generation of the subtracted image. It is important to note that CycleGAN converged in an extremely low number of epochs (14 epochs), which is crucial for model training in terms of computational cost. One significant advantage of CycleGAN is its unsupervised learning manner. CycleGAN works on unpaired data for training. Unlike supervised models such as U-net and Pix2Pix, which require labeled data, CycleGAN uses unlabeled data. This is very useful in medical imaging, where it tends to be challenging and resource-consuming to obtain paired low-energy and contrast-enhanced subtracted images. This model uses CycleGAN, which does not require ground truth and leverages transformation directly from a low-energy image to the subtracted image without the need for supervised representation. Moreover, CycleGAN’s ability to handle the inherent variations in clinical data—such as different patient anatomies and tumor types (or healthy tissues)—contributes to its robustness and generalization capabilities. This unsupervised nature allows the model to adapt to new, unseen data, as evidenced by its superior performance on the test set, which was not used during training. The downside of this approach is that although CycleGAN is capable of producing images from which the high-energy scan has been removed, it inherently depends on the precision of input images and their uniformity. Furthermore, although CycleGAN eliminates the need for dual-energy imaging, it may not capture all the diagnostic information contained in CESM images, especially when very subtle differences between tissue types are essential for a correct diagnosis. Future research is needed to verify its performance under a variety of clinical conditions, as well as to see whether such performance can be generalized beyond those groups of patients where the data have been collected. Conclusion In conclusion, CycleGAN is good for generating pseudo-subtracted images from low-energy mammography data. It can be seen that this method provides comparable results to the conventional two-source imaging technology. The model performs well in producing images with high similarity to subtracted contrast-enhanced ones, offering help in detecting lesions as well as ease of operation. It is this unsupervised nature that makes it especially appropriate for clinical settings where the in most cases labeled data are not available. Moreover, both cost and radiation exposure could be reduced substantially for all patients as part of this technique. Although there are some limitations to the method; for instance, whenever input image quality should go askew, in general, CycleGAN presents a way forward for medical imaging. With testing and still finer refinement, it holds the promise of being applied more widely in clinical practice. Declarations Conflict-of-interest The authors have no relevant financial or non-financial interests to disclose and the authors have no competing interests to declare that are relevant to the content of this article. Funding Statement The authors declare that no funds, grants, or other support were received during the preparation of this manuscript. Author Contributions All authors contributed to the study conception and design. Material preparation, data processing and analysis were performed by Asma Khorshidifar, Ghazal Mostaghel, and Kaveh Dastvareh. The first draft of the manuscript was written by Asma Khorshidifar and Yashar Ahmadyar, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. The idea and conceptualization were carried out by Yashar Ahmadyar. Dr. Rezvan Samimi provided feedback on the study results and revised the manuscript. Declaration of generative AI in scientific writing During the preparation of this work the authors used ChatGPT in order to improve readability and language quality of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication References Harbeck N, Penault-Llorca F, Cortes J, Gnant M, Houssami N, Poortmans P et al (2019) Breast cancer. Nat Rev Dis Primer 5:66 Li L, Roth R, Germaine P, Ren S, Lee M, Hunter K et al (2017) Contrast-enhanced spectral mammography (CESM) versus breast magnetic resonance imaging (MRI): A retrospective comparison in 66 breast lesions. Diagn Interv Imaging 98:113–123 Tang S, Xiang C, Yang Q (2020) The diagnostic performance of CESM and CE-MRI in evaluating the pathological response to neoadjuvant therapy in breast cancer: a systematic review and meta-analysis. Br J Radiol 93:20200301 Sorin V, Sklair-Levy M (2019) Dual-energy contrast-enhanced spectral mammography (CESM) for breast cancer screening. Quant Imaging Med Surg 9:1914–1917 Cheng Z, Wen J, Huang G, Yan J (2021) Applications of artificial intelligence in nuclear medicine image generation. Quant Imaging Med Surg 11:2792–2822 Ouyang J, Chen KT, Duarte Armindo R, Davidzon GA, Hawk KE, Moradi F et al (2024) Predicting FDG-PET Images From Multi‐Contrast MRI Using Deep Learning in Patients With Brain Neoplasms. J Magn Reson Imaging 59:1010–1020 Cai Y, Li M, Liu S, Zhou C (2023) CycleGAN-based image translation from MRI to CT scans. J Phys Conf Ser 2646:012016 McNaughton J, Fernandez J, Holdsworth S, Chong B, Shim V, Wang A (2023) Machine Learning for Medical Image Translation: A Systematic Review. Bioengineering 10:1078 Chen J, Chen S, Wee L, Dekker A, Bermejo I (2023) Deep learning based unpaired image-to-image translation applications for medical physics: a systematic review. Phys Med Biol 68:05TR01 Islam S, Aziz MT, Nabil HR, Jim JR, Mridha MF, Kabir MM et al (2024) Generative Adversarial Networks (GANs) in Medical Imaging: Advancements, Applications, and Challenges. IEEE Access 12:35728–35753 Ferreira A, Li J, Pomykala KL, Kleesiek J, Alves V, Egger J (2024) GAN-based generation of realistic 3D volumetric data: A systematic review and taxonomy. Med Image Anal 93:103100 Department of Data Sciences Harrisburg University of Science & Technology, Harrisburg, Pennsylvania, United States of America, Thakur A, Thakur GK, Department of Data Sciences (2024) Harrisburg University of Science & Technology, Harrisburg, Pennsylvania, United States of America. Developing GANs for Synthetic Medical ImagingData: Enhancing Training and Research. Int J Adv Multidiscip Res. ;11:70–82 Kora P, Ooi CP, Faust O, Raghavendra U, Gudigar A, Chan WY et al (2022) Transfer learning techniques for medical image analysis: A review. Biocybern Biomed Eng 42:79–107 Atasever S, Azginoglu N, Terzi DS, Terzi R (2023) A comprehensive survey of deep learning research on medical image analysis with focus on transfer learning. Clin Imaging 94:18–41 Mukhlif AA, Al-Khateeb B, Mohammed MA (2023) Incorporating a Novel Dual Transfer Learning Approach for Medical Images. Sensors 23:570 Srikantamurthy MM, Rallabandi VPS, Dudekula DB, Natarajan S, Park J (2023) Classification of benign and malignant subtypes of breast cancer histopathology imaging using hybrid CNN-LSTM based transfer learning. BMC Med Imaging 23:19 Dinh P-H, Giang NL (2024) Medical image fusion based on transfer learning techniques and coupled neural P systems. Neural Comput Appl 36:4325–4347 Koul A, Bawa RK, Kumar Y (2024) An Analysis of Deep Transfer Learning-Based Approaches for Prediction and Prognosis of Multiple Respiratory Diseases Using Pulmonary Images. Arch Comput Methods Eng 31:1023–1049 Khajvand N, Ahmadyar Y, Samimi R, Mehrban Q, Kamali-Asl A, Arabi H et al (2024) Whole-Body PET Tumor Lesion Segmentation Using a Transformer-Based Model. 2024 IEEE Nucl Sci Symp NSS Med Imaging Conf MIC Room Temp Semicond Detect Conf RTSD [Internet]. Tampa, FL, USA: IEEE; [cited 2025 Feb 19]. pp. 1–2. Available from: https://ieeexplore.ieee.org/document/10656988/ Zhu J-Y, Park T, Isola P, Efros AA Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks [Internet]. arXiv; 2017 [cited 2025 Feb 19]. Available from: https://arxiv.org/abs/1703.10593 Ali MB, Gu IY-H, Berger MS, Pallud J, Southwell D, Widhalm G et al (2020) Domain Mapping and Deep Learning from Multiple MRI Clinical Datasets for Prediction of Molecular Subtypes in Low Grade Gliomas. Brain Sci 10:463 Additional Declarations The authors declare no competing interests. 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-6147032","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":423441129,"identity":"ff9ecd65-68c0-423a-ae5d-a92803d574b8","order_by":0,"name":"Asma Khorshidifar","email":"","orcid":"","institution":"K.N. Toosi University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Asma","middleName":"","lastName":"Khorshidifar","suffix":""},{"id":423441130,"identity":"e4fe0009-d4ce-489d-9860-ab9fc6f3c1f2","order_by":1,"name":"Ghazal Mostaghel","email":"","orcid":"","institution":"Islamic Azad University of Karaj","correspondingAuthor":false,"prefix":"","firstName":"Ghazal","middleName":"","lastName":"Mostaghel","suffix":""},{"id":423441131,"identity":"8408edbe-8af1-4c0b-8412-964eaf6968ea","order_by":2,"name":"Kaveh Dastvareh","email":"","orcid":"","institution":"Shiraz University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Kaveh","middleName":"","lastName":"Dastvareh","suffix":""},{"id":423441132,"identity":"c2934776-8b68-421f-8a0c-78007a10fea4","order_by":3,"name":"Yashar Ahmadyar","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9ElEQVRIiWNgGAWjYBACgwM8QJINiNn7Hz4AUjx8xGvhOcNsAKLYCGmRbIBpkchhk2CAsvECfvazRzf8KDssbz4j91jl1xw7GTYG5oePbuDRwsaTl3az59xhwzln3qXdlt2WDHQYm7FxDj4tDDlmN3jbDjPOYE8wuy25jRmohYdNGq8W/jdmN/+2HbafwZBgViy5rZ4ILRI5ZreBtiTO4MgxY/y47TAxWt6Y3ZY5l548g+dYsjTjtuM8bMyE/MKfY3bzTZm17Qz25oMff26rtudnb374GJ8WKGgGk8w8YJKwchCoA5OMP4hTPQpGwSgYBSMMAABD40gH8L1GwQAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0003-3966-0164","institution":"Shahid Beheshti University","correspondingAuthor":true,"prefix":"","firstName":"Yashar","middleName":"","lastName":"Ahmadyar","suffix":""},{"id":423441133,"identity":"4bbe75a3-e286-4000-a442-f2c80c5a6077","order_by":4,"name":"Rezvan Samimi","email":"","orcid":"","institution":"Shahid Beheshti University","correspondingAuthor":false,"prefix":"","firstName":"Rezvan","middleName":"","lastName":"Samimi","suffix":""}],"badges":[],"createdAt":"2025-03-03 14:20:15","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6147032/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6147032/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":79256756,"identity":"3abc2a2a-ea1c-43da-8357-a3637b233a66","added_by":"auto","created_at":"2025-03-26 09:01:45","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":122867,"visible":true,"origin":"","legend":"\u003cp\u003eCycleGAN workflow for subtracted image generation.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6147032/v1/e826cfbd9efba26881884488.png"},{"id":79256755,"identity":"40283214-c726-40e9-b9c7-696ae5413f5b","added_by":"auto","created_at":"2025-03-26 09:01:45","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":81968,"visible":true,"origin":"","legend":"\u003cp\u003eCycleGAN architecture (Ali et al. [21]).\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6147032/v1/78f082ee51cfd761639d9a72.png"},{"id":79256761,"identity":"47ff794b-fdfb-4a60-8ac5-fa939199c25f","added_by":"auto","created_at":"2025-03-26 09:01:45","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":313020,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of 2 different patients. From left; input (low energy image), output (virtual subtracted image), and ground truth (original subtracted image)\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6147032/v1/25acfb392991eaadf28a18d3.png"},{"id":79260732,"identity":"d215f05a-120c-42e3-89da-9681a82348ca","added_by":"auto","created_at":"2025-03-26 09:25:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1005066,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6147032/v1/397b528c-b95d-4b68-a84b-fed138885134.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eGenerating Pseudo-Subtracted Image in Dual-Energy Contrast-Enhanced Spectral Mammography Using Transfer Learning\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eBreast cancer is the most common cancer in women. In the early stages, non-metastatic cases are curable in 70\u0026ndash;80% of patients; however, advanced breast cancer with metastases is currently incurable. The disease is molecularly diverse, with features like HER2 activation, hormone receptor positivity, and BRCA mutations influencing treatment. The treatment involves a combination of surgery, radiation therapy, systemic therapies such as endocrine therapy, chemotherapy, anti-HER2 treatment, and immunotherapy [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDual-energy Contrast Enhanced Spectral Mammography (CESM) is a reliable tool for breast cancer diagnosis. The studies have shown that CESM sensitivity is comparable to Breast Magnetic Resonance Imaging (BMRI) in detecting breast cancer. It also has a significantly shorter exam time which has made it an accessible alternative to BMRI as well as an invaluable method in breast cancer detection and staging. Studies have found that CESM offers equal specificity and superior sensitivity in comparison to MRI, and excellent performance in diagnosing the pathological response of breast cancer to neoadjuvant chemotherapy (NAC). Additionally, CESM is more affordable and better tolerated by patients; hence it has become a promising option for evaluating tumors in the breasts [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. CESM in an imaging technique to achieve both anatomical and functional information of the breasts, by capturing two images per view at different energy levels. The first scan is a low-energy image which shows breast structure exactly similar to standard 2D mammography. The second scan is a subtracted image that highlights areas of contrast uptake that may indicate malignant neovascularization [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. However, it is important to note that having two scans per view in CESM increases the radiation dose to the patient and can be time-consuming, which may limit its practical application. Furthermore, the contrast agent used in CESM can potentially harm patients by causing allergic reactions or other adverse effects.\u003c/p\u003e \u003cp\u003eReplacing two scans per view with one can effectively decrease the radiation dose, reduce the risk of allergic reactions to the contrast agent, and shorten the acquisition time. By leveraging deep neural networks, it is possible to virtually generate the subtracted images using only the low-energy scans [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Ouyang et al. proposed a deep convolutional model to generate virtual PET images for brain tumor detection to tackle the challenges of 18F-FDG PET such as cost, radiation exposure, and accessibility. In this retrospective study, convolutional neural networks were trained using multiple MRI sequences to produce synthesized PET images. The study shows the potential of deep learning to provide reliable, diagnostic-quality PET images from MRI, results in a cost-effective and non-invasive alternative for brain neoplasm assessment [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Cai presents a CycleGAN-based method for translating unpaired CT and MRI brain scans for a higher quality tumor diagnosis. The model trained on the Kaggle dataset, and the model demonstrated effective image translation with high quality. CycleGAN\u0026rsquo;s strengths lie in its ability to handle unpaired data with potential for broader applications in medical image analysis [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Moreover, several review articles have shown that image-to-image translation has been successfully applied across various modalities, such as PET, SPECT, MRI, CT, and mammography that facilitate effective cross-modality translation for improved diagnosis and analysis [\u003cspan additionalcitationids=\"CR9 CR10 CR11\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe studies have shown transfer learning is a fast and effective tool in different task of medical imaging [\u003cspan additionalcitationids=\"CR14 CR15 CR16 CR17 CR18\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. In this work, we propose a transfer learning approach to generate subtracted mammography images from low-energy mammography scans.\u003c/p\u003e"},{"header":"Material and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003ea. Dataset and preprocessing\u003c/h2\u003e \u003cp\u003eThe Categorized Digital Database for Low-energy and Subtracted Contrast-Enhanced Spectral Mammography images (CDD-CESM), used in this study, is publicly available. It comprises a total of 1003 low-energy images and 1003 subtracted images (available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.cancerimagingarchive.net/collection/cdd-cesm/\u003c/span\u003e\u003cspan address=\"https://www.cancerimagingarchive.net/collection/cdd-cesm/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). In this dataset, patients received a low-osmolar iodinated contrast agent at a dose of 1.5 ml/kg. Two minutes after administration, the craniocaudal (CC) and mediolateral-oblique (MLO) views were obtained. The first scan is a low-energy exposure with peak kV values from 26 to 31 kV, and the other is a high-energy exposure from 45 to 49 kV. CC and MLO views were obtained and then recombined and subtracted to diminish the background parenchyma after the contrast agent had been injected.\u003c/p\u003e \u003cp\u003eIn the preprocessing part, images were standardized by resizing them to a uniform 256x256 dimension. Furthermore, augmentation techniques were applied, using different rotations, which increased the number of images to 7,600 images. The images were split into a training and blind-testing dataset in a 70%-30% ratio.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eb. Model Architecture\u003c/h3\u003e\n\u003cp\u003eThe Cycle-Consistent Generative Adversarial Network (CycleGAN) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] was employed for the image translation task. Even though image data would seem relatively simple to obtain, in some scenarios, the same image would need to be translated into two different target domains; hence, CycleGAN can be used for this as it does not rely on paired images for training. This is particularly beneficial when paired datasets are not possible, which is frequently the case in medical imaging.\u003c/p\u003e \u003cp\u003eThe main design of CycleGAN is bidirectionality, such that two classes can be transformed, in this instance, from low-energy mammography images to subtracted mammography images, and the opposite way. The model will comprise two generators and two discriminators. The generators are trained to map images from one domain to another (low energy to subtracted and back), and the discriminators are trained to distinguish between real and generated images in each domain This method aids in making sure that the transformation looks similar on both domains (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThere are two main losses that are used while training CycleGAN\u0026mdash;adversarial loss and cycle consistency loss. The adversarial loss is used to promote generated images to be indistinguishable from real images, according to the discriminator. Cycle consistency loss is fundamental to guaranteeing that the transformation is bidirectional, which requires the transformation of an image from domain A (low-energy) to domain B (subtracted) back to domain A to resemble as closely as possible the original input image. This loss guarantees that the network does sensible transformations, i.e., the images keep the meaningful content across transformations. The architecture is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eFor this study, the CycleGAN model was trained using the Adam optimizer with a learning rate of 0.0002, batch size of 1, for 14 epochs. Also, the Pix2Pix, U-Net, and ResNet18 models were trained and tested under the same conditions. The study was conducted using Google Colab's NVIDIA A100 GPU with 40GB of memory, 6912 CUDA cores, and a memory bandwidth of 1.6 TB/s. This high-performance hardware enabled efficient deep learning model training and image processing which provides the computational power necessary to complete the study.\u003c/p\u003e\n\u003ch3\u003ec. Evaluation Strategy\u003c/h3\u003e\n\u003cp\u003eEvaluation of the model\u0026rsquo;s performance was conducted using the following metrics:\u003c/p\u003e\u003cp\u003e1. Mean Squared Error (MSE)\u003c/p\u003e \u003cp\u003eMean Squared Error (MSE) MSE measures the average squared difference between corresponding pixels of the original and generated images.\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:MSE\\:=\\:\\left(\\frac{1}{N}\\right)*\\:sum\\left({\\left({I}_{i}-\\:{I}_{i}^{{\\prime\\:}}\\right)}^{2}\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eN is the total number of pixels.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{I}_{i}\\)\u003c/span\u003e \u003c/span\u003e is the pixel value of the original image at position i.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{I}_{i}^{{\\prime\\:}}\\)\u003c/span\u003e \u003c/span\u003e is the pixel value of the generated image at position i.\u003c/p\u003e\u003cp\u003e2. Peak Signal-to-Noise Ratio (PSNR)\u003c/p\u003e\u003cp\u003ePSNR indicates the ratio between the maximum possible pixel value and the noise level.\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:PSNR\\:=\\:20\\:*\\:log10(MAX\\_I)\\:-\\:10\\:*\\:log10\\left(MSE\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003cem\u003eMAX_I\u003c/em\u003e is the maximum possible pixel value of the image.\u003c/p\u003e \u003cp\u003e3. Structural Similarity Index (SSIM)\u003c/p\u003e\u003cp\u003eSSIM measures the perceptual difference between the original and generated images, focusing on changes in structural information. SSIM values range between \u0026minus;\u0026thinsp;1 and 1, where 1 indicates perfect similarity.\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:SSIM\\:=\\frac{\\left[\\left(2\\:*\\:{\\mu\\:}_{I}*\\:{\\mu\\:}_{I}^{{\\prime\\:}}+\\:C1\\right)*\\:\\left(2\\:*\\:{\\sigma\\:}_{II}^{{\\prime\\:}}+\\:C2\\right)\\right]}{\\left[\\left({\\mu\\:}_{I}^{2}+\\:{\\mu\\:}_{I}^{{\\prime\\:}2}+\\:C1\\right)*\\:\\left({\\sigma\\:}_{I}^{2}+\\:{\\sigma\\:}_{I}^{{\\prime\\:}2}+\\:C2\\right)\\right]}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{{\\mu\\:}}_{\\text{I}}\\text{a}\\text{n}\\text{d}\\:{{\\mu\\:}}_{\\text{I}}^{{\\prime\\:}}\\)\u003c/span\u003e \u003c/span\u003e are the average pixel values of the original and generated images.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{{\\sigma\\:}}_{\\text{I}}^{2}\\text{a}\\text{n}\\text{d}\\:{{\\sigma\\:}}_{\\text{I}}^{{\\prime\\:}2}\\)\u003c/span\u003e \u003c/span\u003e are the variances of the original and generated images.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:\\sigma\\:\\_II{\\prime\\:}\\)\u003c/span\u003e \u003c/span\u003e is the covariance between the original and generated images.\u003c/p\u003e \u003cp\u003eC1 and C2 are small constants to stabilize the division.\u003c/p\u003e \u003cp\u003e4. Mean Absolute Error (MAE)\u003c/p\u003e \u003cp\u003eMAE measures the average absolute difference between corresponding pixels of the original and generated images\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:MAE\\:=\\:\\left(\\frac{1}{N}\\right)*\\:sum\\left(\\left|{I}_{i}-\\:{I}_{i}^{{\\prime\\:}}\\right|\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eN is the total number of pixels.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{I}_{i}\\)\u003c/span\u003e \u003c/span\u003e is the pixel value of the original image at position i.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{I}_{i}^{{\\prime\\:}}\\)\u003c/span\u003e \u003c/span\u003e is the pixel value of the generated image at position i.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThe CycleGAN-generated subtracted images exhibited robust structural and visual similarity to CESM-derived images. These images successfully excluded extraneous features from low-energy images and increased contrast, which helped with the recognition of malignant lesions.\u003c/p\u003e \u003cp\u003eFigures 3 contains sample images including the predicted subtracted images, the low-energy images, and the CESM-generated images\u003c/p\u003e\u003cp\u003eThe metrics for the 1140 unseen test samples are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\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\u003eEvaluation metrics for different models.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePSNR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSSIM\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCycleGAN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e17.962\u003c/b\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e \u003cb\u003e2.158\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.085\u003c/b\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e \u003cb\u003e0.003\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.007\u003c/b\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e \u003cb\u003e0.0003\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.961\u003c/b\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e \u003cb\u003e0.012\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eU-net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.962 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.898\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.285 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.081 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.0010\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.793 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.006\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePix2Pix\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15.962 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.235\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.135 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.018 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.0002\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.833 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.011\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet 18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.367 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.015\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.452 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.026\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.204 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.0041\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.683 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pm\\:\\)\u003c/span\u003e\u003c/span\u003e 0.042\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe results of this study show that CycleGAN may have utility as a possible technique for the generation of virtual subtracted images similar to CESM. In particular, CycleGAN assists with the detection of cancer lesions by diminishing clutter from low-energy input images and enhancing contrast. This means that dual-energy images are not required to generate high-quality subtracted images because CycleGAN is able to provide the end subtracted state given just a low-energy image as input. This enables imaging with minimal optimization and virtually no data collection infrastructure required to create dual-energy data, resulting in cost and practical advantages. This approach minimizes both the hardware requirements and operational costs when obtaining dual-energy images while decreasing the patient\u0026rsquo;s radiation exposure\u0026mdash;an important consideration in medical imaging.\u003c/p\u003e \u003cp\u003eCycleGAN is evaluated on 1140 unseen samples across multiple metrics. CycleGAN outperformed U-net, Pix2Pix, and ResNet18 across all evaluation metrics, notably for structural similarity and peak signal-to-noise ratio, which are key qualities in ensuring the generated images are similar in visual appearance and structure. CycleGAN has a 0.961 structural similarity index score, which shows that its structural similarity to ground truth images is relatively high and that its key features necessary for accurate lesion detection are preserved. PSNR is 17.962\u0026thinsp;\u0026plusmn;\u0026thinsp;2.158, which further demonstrates that CycleGAN is highly capable of image generation. U-net, Pix2Pix, and ResNet18 produced markedly lower peak signal-to-noise ratio and structural similarity index scores, indicating that the models perform poorly in retaining the structural nuances necessary for the appropriate generation of the subtracted image. It is important to note that CycleGAN converged in an extremely low number of epochs (14 epochs), which is crucial for model training in terms of computational cost.\u003c/p\u003e \u003cp\u003eOne significant advantage of CycleGAN is its unsupervised learning manner. CycleGAN works on unpaired data for training. Unlike supervised models such as U-net and Pix2Pix, which require labeled data, CycleGAN uses unlabeled data. This is very useful in medical imaging, where it tends to be challenging and resource-consuming to obtain paired low-energy and contrast-enhanced subtracted images. This model uses CycleGAN, which does not require ground truth and leverages transformation directly from a low-energy image to the subtracted image without the need for supervised representation. Moreover, CycleGAN\u0026rsquo;s ability to handle the inherent variations in clinical data\u0026mdash;such as different patient anatomies and tumor types (or healthy tissues)\u0026mdash;contributes to its robustness and generalization capabilities. This unsupervised nature allows the model to adapt to new, unseen data, as evidenced by its superior performance on the test set, which was not used during training.\u003c/p\u003e \u003cp\u003eThe downside of this approach is that although CycleGAN is capable of producing images from which the high-energy scan has been removed, it inherently depends on the precision of input images and their uniformity. Furthermore, although CycleGAN eliminates the need for dual-energy imaging, it may not capture all the diagnostic information contained in CESM images, especially when very subtle differences between tissue types are essential for a correct diagnosis. Future research is needed to verify its performance under a variety of clinical conditions, as well as to see whether such performance can be generalized beyond those groups of patients where the data have been collected.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, CycleGAN is good for generating pseudo-subtracted images from low-energy mammography data. It can be seen that this method provides comparable results to the conventional two-source imaging technology. The model performs well in producing images with high similarity to subtracted contrast-enhanced ones, offering help in detecting lesions as well as ease of operation. It is this unsupervised nature that makes it especially appropriate for clinical settings where the in most cases labeled data are not available. Moreover, both cost and radiation exposure could be reduced substantially for all patients as part of this technique. Although there are some limitations to the method; for instance, whenever input image quality should go askew, in general, CycleGAN presents a way forward for medical imaging. With testing and still finer refinement, it holds the promise of being applied more widely in clinical practice.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflict-of-interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose and the authors have no competing interests to declare that are relevant to the content of this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that no funds, grants, or other support were received during the preparation of this manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the study conception and design. Material preparation, data processing and analysis were performed by Asma Khorshidifar, Ghazal Mostaghel, and Kaveh Dastvareh. The first draft of the manuscript was written by Asma Khorshidifar and Yashar Ahmadyar, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. The idea and conceptualization were carried out by Yashar Ahmadyar. Dr. Rezvan Samimi provided feedback on the study results and revised the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of generative AI in scientific writing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this work the authors used ChatGPT in order to improve readability and language quality of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHarbeck N, Penault-Llorca F, Cortes J, Gnant M, Houssami N, Poortmans P et al (2019) Breast cancer. Nat Rev Dis Primer 5:66\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi L, Roth R, Germaine P, Ren S, Lee M, Hunter K et al (2017) Contrast-enhanced spectral mammography (CESM) versus breast magnetic resonance imaging (MRI): A retrospective comparison in 66 breast lesions. Diagn Interv Imaging 98:113\u0026ndash;123\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang S, Xiang C, Yang Q (2020) The diagnostic performance of CESM and CE-MRI in evaluating the pathological response to neoadjuvant therapy in breast cancer: a systematic review and meta-analysis. Br J Radiol 93:20200301\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSorin V, Sklair-Levy M (2019) Dual-energy contrast-enhanced spectral mammography (CESM) for breast cancer screening. Quant Imaging Med Surg 9:1914\u0026ndash;1917\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheng Z, Wen J, Huang G, Yan J (2021) Applications of artificial intelligence in nuclear medicine image generation. Quant Imaging Med Surg 11:2792\u0026ndash;2822\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOuyang J, Chen KT, Duarte Armindo R, Davidzon GA, Hawk KE, Moradi F et al (2024) Predicting FDG-PET Images From Multi‐Contrast MRI Using Deep Learning in Patients With Brain Neoplasms. J Magn Reson Imaging 59:1010\u0026ndash;1020\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCai Y, Li M, Liu S, Zhou C (2023) CycleGAN-based image translation from MRI to CT scans. J Phys Conf Ser 2646:012016\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcNaughton J, Fernandez J, Holdsworth S, Chong B, Shim V, Wang A (2023) Machine Learning for Medical Image Translation: A Systematic Review. Bioengineering 10:1078\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen J, Chen S, Wee L, Dekker A, Bermejo I (2023) Deep learning based unpaired image-to-image translation applications for medical physics: a systematic review. Phys Med Biol 68:05TR01\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIslam S, Aziz MT, Nabil HR, Jim JR, Mridha MF, Kabir MM et al (2024) Generative Adversarial Networks (GANs) in Medical Imaging: Advancements, Applications, and Challenges. IEEE Access 12:35728\u0026ndash;35753\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFerreira A, Li J, Pomykala KL, Kleesiek J, Alves V, Egger J (2024) GAN-based generation of realistic 3D volumetric data: A systematic review and taxonomy. Med Image Anal 93:103100\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDepartment of Data Sciences Harrisburg University of Science \u0026amp; Technology, Harrisburg, Pennsylvania, United States of America, Thakur A, Thakur GK, Department of Data Sciences (2024) Harrisburg University of Science \u0026amp; Technology, Harrisburg, Pennsylvania, United States of America. Developing GANs for Synthetic Medical ImagingData: Enhancing Training and Research. Int J Adv Multidiscip Res. ;11:70\u0026ndash;82\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKora P, Ooi CP, Faust O, Raghavendra U, Gudigar A, Chan WY et al (2022) Transfer learning techniques for medical image analysis: A review. Biocybern Biomed Eng 42:79\u0026ndash;107\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAtasever S, Azginoglu N, Terzi DS, Terzi R (2023) A comprehensive survey of deep learning research on medical image analysis with focus on transfer learning. Clin Imaging 94:18\u0026ndash;41\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMukhlif AA, Al-Khateeb B, Mohammed MA (2023) Incorporating a Novel Dual Transfer Learning Approach for Medical Images. Sensors 23:570\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSrikantamurthy MM, Rallabandi VPS, Dudekula DB, Natarajan S, Park J (2023) Classification of benign and malignant subtypes of breast cancer histopathology imaging using hybrid CNN-LSTM based transfer learning. BMC Med Imaging 23:19\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDinh P-H, Giang NL (2024) Medical image fusion based on transfer learning techniques and coupled neural P systems. Neural Comput Appl 36:4325\u0026ndash;4347\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoul A, Bawa RK, Kumar Y (2024) An Analysis of Deep Transfer Learning-Based Approaches for Prediction and Prognosis of Multiple Respiratory Diseases Using Pulmonary Images. Arch Comput Methods Eng 31:1023\u0026ndash;1049\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhajvand N, Ahmadyar Y, Samimi R, Mehrban Q, Kamali-Asl A, Arabi H et al (2024) Whole-Body PET Tumor Lesion Segmentation Using a Transformer-Based Model. 2024 IEEE Nucl Sci Symp NSS Med Imaging Conf MIC Room Temp Semicond Detect Conf RTSD [Internet]. Tampa, FL, USA: IEEE; [cited 2025 Feb 19]. pp. 1\u0026ndash;2. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ieeexplore.ieee.org/document/10656988/\u003c/span\u003e\u003cspan address=\"https://ieeexplore.ieee.org/document/10656988/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu J-Y, Park T, Isola P, Efros AA Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks [Internet]. arXiv; 2017 [cited 2025 Feb 19]. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/abs/1703.10593\u003c/span\u003e\u003cspan address=\"https://arxiv.org/abs/1703.10593\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAli MB, Gu IY-H, Berger MS, Pallud J, Southwell D, Widhalm G et al (2020) Domain Mapping and Deep Learning from Multiple MRI Clinical Datasets for Prediction of Molecular Subtypes in Low Grade Gliomas. Brain Sci 10:463\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Shahid Beheshti University","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":"CESM, Deep learning, Breast cancer","lastPublishedDoi":"10.21203/rs.3.rs-6147032/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6147032/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eDual-energy contrast-enhanced spectral mammography (CESM) enhances breast cancer detection but increases radiation exposure, especially for high-risk patients like BRCA1 mutation carriers. Additionally, the dual-energy acquisition process can be time-consuming. This study uses deep learning to convert low-energy images into subtracted images, reducing radiation and contrast-related risks, while also addressing the time consumption challenge of the traditional CESM procedure.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eThe study utilized the Categorized Digital Database for Low-energy and Subtracted Contrast-Enhanced Spectral Mammography Images (CDD-CESM), which contains 7600 image pairs after augmentation. The dataset was divided into 70% for training and 30% for testing. CycleGAN's performance was evaluated and compared against U-Net, Pix2Pix, and ResNet18. Key metrics for comparison included Structural Similarity Index and Peak Signal-to-Noise Ratio. The models were tested for their ability to generate high-quality subtracted images without the need for paired training data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eCycleGAN outperformed U-Net, Pix2Pix, and ResNet18 in generating pseudo-subtracted images. The SSIM score of 0.961, close to that of real subtracted images, indicates that CycleGAN successfully preserves structural details. Additionally, CycleGAN achieved this performance at a lower computational cost and without the need for paired data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eCycleGAN effectively generates pseudo-subtracted images from low-energy mammography data, presenting a viable alternative to dual-energy imaging. This method has the potential to reduce the need for additional imaging, minimize radiation exposure, and simplify imaging procedures. The high SSIM score highlights CycleGAN's ability to maintain strong structural similarities in the generated images, making it a promising tool for detecting lesions in mammography.\u003c/p\u003e","manuscriptTitle":"Generating Pseudo-Subtracted Image in Dual-Energy Contrast-Enhanced Spectral Mammography Using Transfer Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-03-26 09:01:40","doi":"10.21203/rs.3.rs-6147032/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":"aaf93433-7b23-4827-869f-dce16cf1455d","owner":[],"postedDate":"March 26th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-03-26T09:01:40+00:00","versionOfRecord":[],"versionCreatedAt":"2025-03-26 09:01:40","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6147032","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6147032","identity":"rs-6147032","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}