Enhancing Image Quality of Low-Dose Dental CBCT Using Residual Encoder- Decoder Convolutional Neural Network (RED-CNN): A Comparative Study with Non-Local Means Denoising

Enhancing Image Quality of Low-Dose Dental CBCT Using Residual Encoder- Decoder Convolutional Neural Network (RED-CNN): A Comparative Study with Non-Local Means Denoising | 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 Enhancing Image Quality of Low-Dose Dental CBCT Using Residual Encoder- Decoder Convolutional Neural Network (RED-CNN): A Comparative Study with Non-Local Means Denoising Aprizka Smartalova Syahda, Matthew Gregorius, Syahril Siregar, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8930014/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 7 You are reading this latest preprint version Abstract Objective To evaluate the effectiveness of Residual Encoder-Decoder Convolutional Neural Network (RED-CNN) in reducing noise and improving image quality in low-dose dental Cone-Beam Computed Tomography (CBCT), and to compare its performance with the conventional Non-Local Means (NLM) denoising algorithm. Methods A female head RANDO phantom was scanned using a dental CBCT system with high-dose protocol (90 kV, 10 mA) to obtain ground-truth images and a low-dose protocol (70 kV, 1 mA) to generate noisy datasets. The RED-CNN model was trained and validated on paired low-dose and high-dose images, and tested on unseen data to assess generalization performance. Quantitative evaluation included Signal Difference-to-Noise Ratio (SDNR), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), computational performance, and dose-reduction assessment using the Dose Area Product (DAP). Results Both methods reduced noise and enhanced the quality of low-dose dental CBCT images; however, RED-CNN consistently outperformed NLM across all quantitative metrics. RED-CNN achieved SDNR > 25, PSNR > 30 dB, and SSIM of 0.73, demonstrating improved noise suppression and preservation of anatomical structures, and also provided faster GPU-based inference. Dose analysis showed that the low-dose protocol reduced exposure by 94% while maintaining acceptable diagnostic quality. Conclusions These findings showed that low-dose dental CBCT may achieve clinically acceptable image quality when enhanced using RED-CNN, as the method effectively suppresses noise while preserving essential anatomical detail. CBCT low-dose imaging image denoising RED-CNN NLM Figures Figure 1 Figure 2 Figure 3 Introduction Cone-beam computed tomography (CBCT) is a three-dimensional imaging modality that provides detailed visualisation of dentomaxillofacial structures, including the jaws and surrounding anatomy 1 , 2 . Compared with conventional two-dimensional imaging techniques such as periapical and panoramic radiography, CBCT offers improved geometric accuracy and spatial resolution with reduced image superimposition. Furthermore, CBCT delivers a substantially lower radiation dose than conventional multidetector computed tomography (MDCT), making it a valuable imaging tool in dental and maxillofacial applications. The control of the field of view (FOV) directly affects radiation exposure and voxel size, which in turn influence image resolution and diagnostic quality 3 . As a radiographic technique utilising ionising radiation, CBCT imaging should be employed in accordance with radiation protection principles, including limiting exposure to the region of interest (ROI) and following the ALARA (As Low As Reasonably Achievable) principle, to minimise cumulative radiation dose and the associated risks of cancer and tissue effects 4 . However, reducing radiation exposure inevitably increases image noise 5 , which can degrade image quality and affect diagnostic accuracy. To maintain diagnostic reliability, noise reduction methods are required to suppress unwanted artefacts while preserving important anatomical details. Several image denoising methods have been proposed for low-dose imaging. In recent years, deep learning-based reconstruction (DLR) and post-processing approaches have demonstrated superior performance by learning noise patterns and image structures using convolutional neural networks (CNNs). Among these approaches, the residual encoder–decoder convolutional neural network (RED-CNN) proposed by Chen et al. 6 combines an encoder–decoder framework with residual skip connections, enabling effective noise suppression while maintaining fine structural details in low-dose CT imaging. Previous studies have reported that RED-CNN improves image quality in low-dose CT and 3D rotational angiography (3DRA), with significant increases in signal-to-noise ratio (SNR), PSNR, and SSIM 7 , 8 . These findings indicate that RED-CNN can achieve effective noise suppression while maintaining diagnostic fidelity, indicating its potential suitability for low-dose dental CBCT imaging. Therefore, this study aims to implement and evaluate the performance of RED-CNN for noise suppression in low-dose dental CBCT images. A quantitative evaluation will be conducted using the SDNR, PSNR, SSIM, computational performance, and estimated radiation dose-reduction efficiency to assess the model’s denoising capability. The performance of RED-CNN will be compared with the conventional non-local means (NLM) denoising algorithm 9 . The dataset used in this study comprises ten low-dose and ten standard-dose dental CBCT images acquired at the Universitas Indonesia Hospital (RSUI) using a female head RANDO anthropomorphic phantom, ensuring consistency and reproducibility by eliminating inter-individual anatomical variability. Consequently, this research focuses on quantitative algorithmic evaluation rather than clinical validation in patients. Methods and materials Image Acquisition CBCT imaging was performed using a 3D Accuitomo system (Model MCT-1, J. Morita MFG. Corp., Kyoto, Japan). The system was equipped with a high-resolution flat-panel detector (FPD) capable of volumetric three-dimensional acquisition with isotropic voxels 10 . Image reconstruction and visualisation were carried out using the manufacturer’s i-Dixel software (J. Morita). Imaging parameters were standardised between the low-dose and high-dose protocols to ensure geometric consistency and reproducibility across all scans. The imaging object was a RANDO anthropomorphic phantom (The Phantom Laboratory, USA), which represented human head and neck anatomy to ensure consistency, reproducibility, and quantitative validity during evaluation 11 , 12 . Two exposure protocols were applied to simulate clinical conditions. The low-dose protocol, used as the noisy input, employed 70 kV and 1 mA, whereas the high-dose protocol, used as the ground-truth reference, employed 90 kV and 10 mA. Both protocols used an identical FOV (140 × 100 mm) and scan time (17.5 s) to maintain consistent imaging geometry. Scans were performed in standard imaging mode (CT mode, 360° rotation), with the phantom precisely positioned to ensure spatial alignment between low-dose and high-dose datasets. Each scan yielded 201 axial slices in 16-bit greyscale DICOM format, with an in-plane resolution of 561 × 561 pixels. Ethics and compliance This study did not involve human participants. All image data were acquired using a female head RANDO anthropomorphic phantom; therefore, ethical approval and informed consent were not required. AI-assisted language-editing tools were used solely to improve grammatical accuracy and clarity. These tools did not contribute to data analysis, model development, and interpretation of results or conclusions. The authors take full responsibility for the scientific content of the manuscript. Data preprocessing From each CBCT acquisition, 201 paired low-dose and high-dose axial slices were obtained. Ten paired volumes were used for RED-CNN training and validation. In comparison, an additional 10 low-dose volumes that had not been included previously were used for inference or testing to assess generalisation performance. Before processing with the RED-CNN algorithm, DICOM files were imported using the pydicom library and converted to Hounsfield Units (HU) using the RescaleSlope and RescaleIntercept parameters 13 . Images were adjusted to fit an intensity range of 0–1 using the HU values. The HU values were limited to the hard-tissue window between − 500 and + 1500, which is the standard range used in CT imaging for seeing cortical and trabecular structures. This range covers air, soft tissues, and high-density areas in the dentomaxillofacial region 14 . Each image was cropped into 512 x 512 pixels to remove air regions and fully cover the RANDO phantom region to reduce computational load and increase training relevance. To improve model generalization, data augmentation was applied through rotation and horizontal/vertical flipping. Augmented images preserved anatomical integrity while producing variations in pixel distributions useful for model learning. To increase the number of training samples, patch extraction was performed, dividing images into overlapping 64x64 pixels using a stride of 8. The denoising model was based on the RED-CNN architecture proposed by Chen et al. 6 . The network was implemented using TensorFlow/Keras framework in Python. The input to the network consisted of 2D patches of size 64 x 64 pixels with a single channel (greyscale), corresponding to low-dose dental CBCT images. The RED-CNN architecture comprised five convolutional layers in the encoder path and five corresponding deconvolutional (transposed convolutional) layers in the decoder path. Each convolutional and deconvolutional layer used 96 filters with a kernel size of 5×5 and “same” padding to preserve spatial dimensions. Rectified linear unit (ReLU) activation functions were applied after each convolutional layer. Feature maps from each encoder layer were stored and later added element-wise to the corresponding decoder layer outputs via local residual (skip) connection. A final transposed convolutional layer with a single filter produced the residual, which was then added to the original input patch via a global residual connection. RED-CNN Architecture The model was trained using the Adam optimiser with an initial learning rate of 1 × 10 − 4. A step-decay learning rate schedule was applied, with the learning rate reduced by a factor of 0.991 every 40 epochs 6 , 7 , 15 . The loss function was the MSE between the predicted denoised patches and the corresponding high-dose target patches as expressed in Eq. 1. $$Loss=\frac{1}{N}\sum_{i}^{N}{‖f\left({X}_{i}\right)-{Y}_{i}‖}^{2}$$ where \({Y}_{i}\) and \({X}_{i}\) represents high-dose and low-dose CBCT images. The mapping function \(f\) is used on the low-dose inputs to generate outputs that are as close as possible to the ground truth images 6 . Training and validation data sets were constructed using 80% or 20% splits of low-dose or high-dose patch pairs, as described in the dataset preparation section. To improve computational efficiency, the training data were organized using the tf.data pipeline, with shuffling and mini-batch training. The batch size was set to 16, and prefetching was enabled to optimise GPU utilization. After training, the final RED-CNN model was used to reconstruct whole low-dose CBCT slices by applying the network to each patch and reassembling the outputs into denoised images. The trained RED-CNN model was applied to whole low-dose CBCT slices to produce denoised reconstructions. The model was loaded into TensorFlow/Keras and used to process each slice patch-wise, consistent with the patch extraction strategy used during training. DICOM images were first imported using the pydicom library, converted to Hounsfield Unit (HU) values, and normalized to the range 0–1 using the same clinical used in the training pipeline. To perform reconstruction, each normalized slice was divided into overlapping 64 x 64-pixel patches with a stride of 8. The patches were then input to the RED-CNN model in mini-batches for efficient GPU-accelerated inference. The predicted denoised patches were reassembled into a full-size image using Gaussian-weighted blending to minimise boundary artefacts and ensure smooth patch transitions. After reconstruction, the output image was denormalized back into HU values using the inverse windowing function. The reconstructed HU image retained all original DICOM metadata. It was written back into DICOM format by converting HU values into stored pixel values using the original RescaleSlope and RescaleIntercept. This ensured compatibility with CBCT viewing workstations and enabled direct comparison of low-dose, high-dose, and RED-CNN reconstructed images. Following reconstruction, quantitative evaluation of the denoised images was conducted to assess the performance of the RED-CNN model. The test data comprised low-dose CBCT slices excluded from the training phase. The resulting outputs were compared with the high-dose images as ground truth images. This evaluation determined whether the model could reduce noise and produce high-quality CBCT images while maintaining essential anatomical structures. Comparative method: NLM In addition, a comparison was made between the denoising results obtained with the RED-CNN algorithm and those obtained with NLM, based on the previously obtained evaluation metrics. NLM is a conventional denoising method that can improve the image quality of low-dose dental CBCT. A comparative evaluation of the performance of RED-CNN and NLM was also conducted, focusing on the computational efficiency of both denoising methods for inference/image reconstruction. To apply NLM to dental CBCT images, we need to determine the optimal input settings. We can determine optimal NLM settings by analyzing the PSNR values. These input settings are found through experiments 16 . The input parameters include patch size (5 × 5, 7 × 7, 9 × 9, 11 × 11, 15 × 15), patch distance (5, 7, 9, 11, 15), and h (0.10, 0.12, 0.14, 0.16, 0.18, 0.20, 0.22, 0.24). Evaluation metrics The image quality assessment was based on three primary quantitative metrics, including SDNR, PSNR, dan SSIM 6 , 17 . SDNR measurement is performed by selecting two different regions of interest (ROI), where ROI 1 is for the object area, and ROI 2 is for the background area, as shown in Fig. 1 . This method enabled evaluation of contrast discrimination relative to background noise. In addition to assessing image quality improvement, this study also examined whether the denoising approach could support the use of lower radiation exposures without compromising diagnostic reliability. The radiation output for the CBCT scan was measured using the DAP, which represents the total X-ray energy delivered across the FOV and serves as a standard measure of patient radiation burden in CBCT examinations. DAP is routinely recorded by most CBCT systems and forms the basis for estimating effective dose across different exposure settings, as reported by Ludlow et.al 18 Comparing the DAP values obtained from the high-dose and low-dose protocols allowed us to evaluate whether diagnostic-quality images could still be achieved under substantially reduced radiation levels. Comparing DAP values from the high-dose and low-dose protocols enabled assessment of whether diagnostic-quality images could be maintained under significantly reduced radiation levels. Dose-reduction efficiency was calculated using: $$\%Doseefficiency=\frac{{DAP}_{high}-{DAP}_{low}}{{DAP}_{high}}\times100\%$$ where \({DAP}_{high}\) denotes the radiation output of the high-dose protocol, and \({DAP}_{low}\) denotes the output of the low-dose protocol. Computational performance was also evaluated to determine the practicality of each denoising method in a clinical workflow. Metrics included execution time per slice, GPU/ CPU utilization, and memory consumption, enabling comparison between GPU-accelerated deep learning inference and CPU-based NLM processing. Results The visual quality of the high-dose CBCT image (Fig. 1 A) was markedly superior to that of the low-dose acquisition (Fig. 1 B), which exhibited increased noise and reduced contrast. This observation is consistent with fundamental radiographic principles, whereby reductions in tube voltage and tube current decrease radiation output, but simultaneously increase quantum noise and degrade image quality 19 . Quantitative evaluation demonstrated that the high-dose CBCT images consistently exhibited high SDNR values, as shown in Fig. 2 A. The SDNR range for high-dose images was 12.52–13.34, indicating that the images had good contrast between tissues and minimal noise. In contrast, the low-dose protocol returned SDNR values of 2.33–2.41, reflecting substantial image degradation. According to the Rose model, an SDNR greater than 5 is generally required for reliable visual detection 20 , 21 , supporting the need for denoising enhancement in low-dose CBCT imaging. Application of the RED-CNN model to low-dose CBCT slices resulted in a marked improvement in image quality. SDNR values increased to a range of 27.87–36.44, with a mean SDNR value of 32.87. This increase indicates substantial suppression of background noise and improved signal differentiation. Although SDNR values exceeded those of the high-dose reference images, this increase primarily reflects variance suppression, a known and expected effect of image denoising. Importantly, no loss of anatomical structures was visually observed, indicating that the enhancement remained diagnostically acceptable. The PSNR also improved significantly, as shown in Fig. 2 B. Low-dose images exhibited PSNR values of 17.88–18.20 dB, and RED-CNN denoised images reached 30.10–30.74 dB, corresponding to an improvement of approximately 12 dB. As PSNR values above 30 dB are generally considered indicative of acceptable diagnostic image quality, these findings indicate that RED-CNN-denoised images approximate high-dose image quality to a substantial extent. This result is consistent with findings reported by Chen et al. 6 who observed PSNR improvements of 10 dB to 13 dB in low-dose CT denoising. SSIM analysis provided further evidence of structural preservation. As shown in Fig. 3 A, low-dose CBCT images exhibited SSIM values of approximately 0.30, whereas RED-CNN-denoised images achieved SSIM values ranging from 0.72 to 0.75 (mean 0.73). The results indicate substantial structural similarity to high-dose (ground truth) images and show that RED-CNN better preserves anatomical contours and tissue interfaces than its conventional counterparts. For comparison, the NLM denoising was also applied. Parameter optimisation was performed through empirical testing, and the highest PSNR was achieved with a patch size of 11 × 11, a patch distance of 9, and a filtering parameter \(h\) of 0.14 (Fig. 3 B). NLM improves SDNR to 23.13 ± 1.78 (Fig. 2 A), PSNR to 27.6 ± 0.29 dB (Fig. 2 B), and SSIM to 0.67 ± 0.0093 (Fig. 3 A). Although NLM effectively reduced random noise, visual inspection revealed subtle blurring of fine anatomical details, consistent with its non-parametric averaging behaviour. The results indicate improved noise suppression and structural preservation in RED-CNN-denoised images compared with NLM, particularly in regions with high anatomical frequency. The computational performance of both methods is summarised in Table 1 . RED-CNN required an average processing time of 14.1 s per function call (42.2 s per slice), whereas NLM required 48.1 s per function call (144.1 s per slice)Radiation dose output for each CBCT protocol was assessed using the dose-area product (DAP). The high-dose protocol yielded a DAP of 4300 mGy. cm 2 , whereas the low-dose protocol yielded a DAP of 238 mGy.cm 2 , corresponding to a dose reduction of approximately 94%. Despite this substantial dose reduction, RED-CNN successfully restored low-dose image quality to levels approaching those of high-dose images, as reflected by PSNR and SSIM values within clinically acceptable ranges. These findings demonstrate that deep learning-based denoising can effectively mitigate image quality degradation associated with dose reductions. The enhanced image quality enabled by RED-CNN contributes to clinically compatible visualisation of dental and maxillofacial structures even at the lowest radiation dosages. From a radiation protection perspective, the results support optimisation strategies consistent with the ALARA principle, enabling diagnostically acceptable CBCT imaging at substantially reduced radiation doses. This is especially helpful for repeated imaging procedures and for radiosensitive patient populations, such as children. Discussion This study demonstrates that RED-CNN provides an effective approach for restoring diagnostically acceptable image quality in low-dose dental CBCT imaging. The model consistently improved noise suppression and structural preservation beyond what was achieved with conventional NLM filtering, enabling high-quality visualisation even when the radiation dose was reduced by approximately 94%. These findings have important implications for dose optimisation in dentomaxillofacial imaging. The image quality degradation observed in the raw low-dose images is consistent with established radiographic principles. Reductions in tube voltage or tube current decrease photon fluence, leading to increased quantum noise, reduced contrast resolution, and obscuration of fine anatomical detail. These limitations were clearly reflected in the low SDNR, PSNR, and SSIM values of the unprocessed low-dose images. The improvements achieved using RED-CNN demonstrate the capability of deep-learning models to mitigate these physical limitations by learning the relationship between noise characteristics and underlying image structures from paired training data. In comparison, NLM provided moderate improvements in objective image quality metrics but introduced blurring of high-frequency anatomical structures, which may reduce diagnostic confidence. By contrast, RED-CNN preserved dental and osseous contours while effectively suppressing stochastic noise, supporting its suitability for clinical interpretation. The improved structural fidelity achieved by RED-CNN reflects the advantage of residual learning in retaining edge detail and spatial gradients that are crucial in dental and maxillofacial evaluation. The observed reduction in processing time reflects the contrast between GPU-accelerated feed-forward inference in RED-CNN and the iterative non-local patch search employed by CPU-based NLM. The relatively rapid inference time achieved by RED-CNN suggests potential suitability for integration into time-constrained clinical imaging workflows, where efficient image processing is desirable. From a radiation protection perspective, the most important implication of this study is the demonstrated ability to achieve substantial dose reduction while maintaining diagnostically acceptable image quality. A reduction in radiation dose of approximately 94% directly supports the principle of ALARA and ALADAIP. This capability is especially relevant for paediatric imaging, orthodontic follow-up, implant planning, and other clinical scenarios that require repeated CBCT examinations. The findings suggest that deep learning-based denoising may enable the implementation of ultra-low-dose CBCT protocols without compromising diagnostic utility. Overall, the results indicate that RED-CNN is a promising tool for enabling significant dose reduction in dental CBCT imaging while ensuring high-quality imaging. Future studies should investigate the generalisability of the model across different CBCT systems, anatomical regions, and clinical conditions, as well as its potential integration into vendor software for real-time reconstructions. Conclusion The study demonstrated that the RED-CNN deep learning architecture provides an effective solution for noise suppression and low-dose dental CBCT imaging. The proposed approach substantially improved image quality, with SDNR values exceeding 5 and reaching up to 25, PSNR increasing by more than 10 dB to values above 30 dB, and SSIM rising to approximately 0.7, indicating substantial similarity to high-dose (ground-truth) images. Compared with conventional NLM denoising, RED-CNN consistently outperformed NLM across all quantitative image quality metrics, including SDNR, PSNR, and SSIM. In addition, RED-CNN achieved faster computational performance, demonstrating superior efficiency in both noise suppression and structural preservation. These findings highlight the robustness of deep learning-based approaches relative to traditional patch-based denoising methods. Optimisation of the low-dose CBCT acquisition protocol resulted in an approximately 94% reduction in radiation exposure compared with the high-dose protocol, without modification of other acquisition parameters such as FOV or scan time. Together, these results support the feasibility of combining low-dose CBCT protocols with advanced deep learning denoising to maintain diagnostic utility while adhering to the ALARA principle. Overall, RED-CNN represents a promising and clinically relevant approach for enhancing low-dose CBCT imaging while enabling substantial radiation dose reduction without compromising diagnostic image quality. Declarations Funding acknowledgement This study was conducted without financial support from any public, commercial, or not-for-profit funding agency. All resources utilized in this research were provided by the authors' affiliated institutions. Author Contribution A.S.S. collected all data and wrote the main manuscript, M.G. and S.S. developed the algorithm, L.E.L. composed the research idea and design. All authors reviewed the manuscript. Acknowledgement The authors thank the Radiology Unit of Universitas Indonesia Hospital for providing access to the CBCT imaging facilities used in this study. References Nemtoi A, Czink C, Haba D, Gahleitner A (2013) Cone beam CT: a current overview of devices. Dentomaxillofac Radiol 42(8). 10.1259/DMFR.20120443 Kazimierczak W, Wajer R, Wajer A, Kazimierczak N, Janiszewska-Olszowska J Enhanced Cone-Beam Computed Tomography Imaging through Deep Learning Model Reconstruction: Noise Reduction and Image Quality Optimization in Dental Diagnostics. Published online 2023. 10.21203/rs.3.rs-3650822/v1 Kaasalainen T, Ekholm M, Siiskonen T, Kortesniemi M (2021) Dental cone beam CT: An updated review. Physica Medica Associazione Italiana di Fisica Medica 88:193–217. 10.1016/j.ejmp.2021.07.007 CLEMENT CH, HAMADA N Radiological protection in cone beam 11 computed tomography (CBCT) 12. ICRP PUBLICATION . 1(7):8 Pauwels R (2015) Cone beam CT for dental and maxillofacial imaging: dose matters. Radiat Prot Dosimetry 165(1–4):156–161 Chen H, Zhang Y, Kalra MK et al (2017) Low-Dose CT with a residual encoder-decoder convolutional neural network. IEEE Trans Med Imaging 36(12):2524–2535. 10.1109/TMI.2017.2715284 Cokrokusumo HB, Hariyati I, Lubis LE, Prajitno P, Soejoko DS (2022) Noise Suppression of Computed Tomography (CT) Images Using Residual Encoder-Decoder Convolutional Neural Network (RED-CNN). Atom Indonesia 48(3):171–178. 10.17146/AIJ.2022.1113 Lubis LE, Cokrokusumo HB, Basith RA, Lestariningsih I, Prajitno P, Soejoko DS (2021) Noise reduction of three-dimensional rotational angiography (3DRA) images using residual encoder-decoder convolutional neural network (RED-CNN). In: AIP Conference Proceedings . Vol 2346. American Institute of Physics Inc.; 10.1063/5.0048077 Buades A, Coll B, Morel JM (2011) Non-Local Means Denoising. Image Process Line 1:208–212. 10.5201/ipol.2011.bcm_nlm Mah E, Ritenour ER, Yao H (2020) A review of dental cone-beam CT dose conversion coefficients. Dentomaxillofacial Radiology British Inst Radiology 50(3). 10.1259/dmfr.20200225 Theodorakou C, Walker A, Horner K, Pauwels R, Bogaerts R, Jacobs R (2012) Estimation of paediatric organ and effective doses from dental cone beam CT using anthropomorphic phantoms. Br J Radiol 85(1010):153–160. 10.1259/bjr/19389412 Hörnlund M, Bernhardsson C (2013) ORGAN DOSES DETERMINED USING A RANDO PHANTOM FOR DIFFERENT RADIONUCLIDE DEPOSITIONS AND PHOTON ENERGIES . Vol 11 Magill D, Beckmann N, Felice MA, Yoo T, Luo M, Mupparapu M (2018) Investigation of dental cone-beam CT pixel data and a modified method for conversion to hounsfield unit (HU). Dentomaxillofacial Radiol 47(2). 10.1259/dmfr.20170321 Yu L, Vrieze TJ, Bruesewitz MR et al (2010) Dose and image quality evaluation of a dedicated cone-beam CT system for high-contrast neurologic applications. Am J Roentgenol 194(2). 10.2214/AJR.09.2951 Kingma DP, Ba J, Adam (2014) A Method for Stochastic Optimization. Published online Dec 22, http://arxiv.org/abs/1412.6980 Siregar S, Nagaoka R, Ul Haq I, Saijo Y (2018) Non local means denoising in photoacoustic imaging. Jpn J Appl Phys 57(7). 10.7567/JJAP.57.07LB06 Wang Z, Bovik AC, Sheikh HR, Simoncelli EP (2004) Image quality assessment: From error visibility to structural similarity. IEEE Trans Image Process 13(4):600–612. 10.1109/TIP.2003.819861 Ludlow JB, Timothy R, Walker C et al (2015) Ffective dose of dental CBCT - A meta analysis of published data and additional data for nine CBCT units. Dentomaxillofacial Radiology British Inst Radiology 44(1). 10.1259/dmfr.20140197 Lorenzoni DC, Bolognese AM, Garib DG, Guedes FR, Sant’Anna EF (2012) Cone-beam computed tomography and radiographs in dentistry: Aspects related to radiation dose. Int J Dent . Preprint posted online. 10.1155/2012/813768 Bushberg JT (2012) The Essential Physics of Medical Imaging. Wolters Kluwer Health/Lippincott Williams & Wilkins Monnin P, Gnesin S, Verdun FR, Marshall NW (2019) Generalized SDNR analysis based on signal and noise power. Physica Med 64:10–15. 10.1016/j.ejmp.2019.06.005 Tables Table 1. Computational performance between RED-CNN and NLM Computational Aspect RED-CNN NLMD Execution time per slice ±42,2 s ±144,1 s Mean call ±14,1 s ±48 s Repeatitions 3 3 Execution mode GPU CPU GPU utilization Average 8%, peak 14% 0% Memory usage 75 MB RAM + 3,9-4,0 VRAM GPU 6,5 MB RAM + 0,34 VRAM Computational model Deep Learning Algorithmic, Non-learning Computational efficiency High efficiency Lower efficiency Hardware requirements GPU recommended for optimal performance CPU-only is sufficient, but with a prolonged processing time Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 30 Apr, 2026 Reviews received at journal 08 Apr, 2026 Reviewers agreed at journal 17 Mar, 2026 Reviewers invited by journal 16 Mar, 2026 Editor assigned by journal 28 Feb, 2026 Submission checks completed at journal 28 Feb, 2026 First submitted to journal 20 Feb, 2026 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-8930014","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":607694222,"identity":"cd16fc63-4600-4096-a985-12d536bdc921","order_by":0,"name":"Aprizka Smartalova Syahda","email":"","orcid":"","institution":"Universitas Indonesia","correspondingAuthor":false,"prefix":"","firstName":"Aprizka","middleName":"Smartalova","lastName":"Syahda","suffix":""},{"id":607694223,"identity":"0b1933cb-11cb-4280-852f-f87f90358698","order_by":1,"name":"Matthew Gregorius","email":"","orcid":"","institution":"Universitas Indonesia","correspondingAuthor":false,"prefix":"","firstName":"Matthew","middleName":"","lastName":"Gregorius","suffix":""},{"id":607694224,"identity":"840916e4-5d90-4f26-b3ad-ce091a8244a2","order_by":2,"name":"Syahril Siregar","email":"","orcid":"","institution":"Universitas Indonesia","correspondingAuthor":false,"prefix":"","firstName":"Syahril","middleName":"","lastName":"Siregar","suffix":""},{"id":607694225,"identity":"53c9f8b1-35d8-475c-a81a-a638cfd694ed","order_by":3,"name":"Lukmanda Evan Lubis","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAklEQVRIiWNgGAWjYBACPgY2EMUM5VYwgPkHQGw2hgSsWthQtZwhWQtjG4okDi0Saakbfu6wlmdg7zH8XDivTp6PgcfwAEONHQMfO04tx272nkk3bOA5Yyw9c9thwzYGHoMDDMeSGdh4HuDQkt52g7ftMGODRFqCNO+2AwlsYC1sQCSBy5b0tpt/2w7bN8g/S/7NO6cOquUfPi1px24DbUlskGA+Js3bwAzRwtiGRwvPs7Tbsm3pyW08yceseY4B/cLMVnAgsS+ZB5df+NnTzG6+bbO27Wc/2Hybp6ZOXr69efOHD9/s5OTbsduCsA7OAsURUDEPfvWjYBSMglEwCvABAG5AUNpj8xTHAAAAAElFTkSuQmCC","orcid":"","institution":"Universitas Indonesia","correspondingAuthor":true,"prefix":"","firstName":"Lukmanda","middleName":"Evan","lastName":"Lubis","suffix":""}],"badges":[],"createdAt":"2026-02-21 02:38:03","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8930014/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8930014/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105035698,"identity":"f8c16b84-755a-445b-a41f-c4dcb6537fcb","added_by":"auto","created_at":"2026-03-20 07:26:28","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":435176,"visible":true,"origin":"","legend":"\u003cp\u003eFour axial dental CBCT images of the mandible are arranged in a 2 × 2 layout. Panel A shows a high-dose reference image with two square regions of interest marked in the soft tissue and background areas. Panel B shows a low-dose image with increased image noise. Panel C shows a denoised image using non-local means (NLM) denoising with reduced noise and preserved mandibular structures. Panel D shows a RED-CNN-denoised image with a smoother background appearance and more clearly defined anatomical boundaries.\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8930014/v1/f3286a9e1a5b5cfcf7c86b43.jpg"},{"id":104994049,"identity":"c4b67200-16f2-4dad-8c44-b12fb45decf7","added_by":"auto","created_at":"2026-03-19 15:58:51","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":163705,"visible":true,"origin":"","legend":"\u003cp\u003eTwo quantitative plots are arranged side by side. Panel A shows a bar chart comparing SSIM values for low-dose, NLM-denoised, and RED-CNN-denoised CBCT images. Panel B shows a line graph of PSNR values plotted against the filter parameter h for multiple patch sizes, with a marked peak PSNR value at a specific parameter setting.\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8930014/v1/d38cbfff18906e8738f5a2e7.jpg"},{"id":104994059,"identity":"e20b25f7-4e80-453b-8741-a2bd1ad58776","added_by":"auto","created_at":"2026-03-19 15:58:51","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":180550,"visible":true,"origin":"","legend":"\u003cp\u003eQuantitative comparison of image quality metrics for low-dose dental CBCT images. (A) Structural similarity index (SSIM) values for low-dose, NLMD-denoised, and RED-CNN-denoised images. (B) Peak signal-to-noise ratio (PSNR) as a function of the filter parameter h for different patch sizes, with a fixed patch distance of 9 pixels. The optimal PSNR value is highlighted.\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8930014/v1/b02900d9deb293736a9456b3.jpg"},{"id":105037755,"identity":"982a6087-2026-4a57-a993-ad6adbe2c515","added_by":"auto","created_at":"2026-03-20 07:40:17","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1288846,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8930014/v1/69638735-a06f-4612-b32d-2df070c99d34.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Enhancing Image Quality of Low-Dose Dental CBCT Using Residual Encoder- Decoder Convolutional Neural Network (RED-CNN): A Comparative Study with Non-Local Means Denoising","fulltext":[{"header":"Introduction","content":"\u003cp\u003eCone-beam computed tomography (CBCT) is a three-dimensional imaging modality that provides detailed visualisation of dentomaxillofacial structures, including the jaws and surrounding anatomy\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Compared with conventional two-dimensional imaging techniques such as periapical and panoramic radiography, CBCT offers improved geometric accuracy and spatial resolution with reduced image superimposition. Furthermore, CBCT delivers a substantially lower radiation dose than conventional multidetector computed tomography (MDCT), making it a valuable imaging tool in dental and maxillofacial applications. The control of the field of view (FOV) directly affects radiation exposure and voxel size, which in turn influence image resolution and diagnostic quality\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAs a radiographic technique utilising ionising radiation, CBCT imaging should be employed in accordance with radiation protection principles, including limiting exposure to the region of interest (ROI) and following the ALARA (As Low As Reasonably Achievable) principle, to minimise cumulative radiation dose and the associated risks of cancer and tissue effects\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. However, reducing radiation exposure inevitably increases image noise\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, which can degrade image quality and affect diagnostic accuracy. To maintain diagnostic reliability, noise reduction methods are required to suppress unwanted artefacts while preserving important anatomical details.\u003c/p\u003e \u003cp\u003eSeveral image denoising methods have been proposed for low-dose imaging. In recent years, deep learning-based reconstruction (DLR) and post-processing approaches have demonstrated superior performance by learning noise patterns and image structures using convolutional neural networks (CNNs). Among these approaches, the residual encoder\u0026ndash;decoder convolutional neural network (RED-CNN) proposed by Chen et al.\u003csup\u003e6\u003c/sup\u003e combines an encoder\u0026ndash;decoder framework with residual skip connections, enabling effective noise suppression while maintaining fine structural details in low-dose CT imaging.\u003c/p\u003e \u003cp\u003ePrevious studies have reported that RED-CNN improves image quality in low-dose CT and 3D rotational angiography (3DRA), with significant increases in signal-to-noise ratio (SNR), PSNR, and SSIM\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. These findings indicate that RED-CNN can achieve effective noise suppression while maintaining diagnostic fidelity, indicating its potential suitability for low-dose dental CBCT imaging.\u003c/p\u003e \u003cp\u003eTherefore, this study aims to implement and evaluate the performance of RED-CNN for noise suppression in low-dose dental CBCT images. A quantitative evaluation will be conducted using the SDNR, PSNR, SSIM, computational performance, and estimated radiation dose-reduction efficiency to assess the model\u0026rsquo;s denoising capability. The performance of RED-CNN will be compared with the conventional non-local means (NLM) denoising algorithm\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. The dataset used in this study comprises ten low-dose and ten standard-dose dental CBCT images acquired at the Universitas Indonesia Hospital (RSUI) using a female head RANDO anthropomorphic phantom, ensuring consistency and reproducibility by eliminating inter-individual anatomical variability. Consequently, this research focuses on quantitative algorithmic evaluation rather than clinical validation in patients.\u003c/p\u003e"},{"header":"Methods and materials","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eImage Acquisition\u003c/h2\u003e \u003cp\u003eCBCT imaging was performed using a 3D Accuitomo system (Model MCT-1, J. Morita MFG. Corp., Kyoto, Japan). The system was equipped with a high-resolution flat-panel detector (FPD) capable of volumetric three-dimensional acquisition with isotropic voxels\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Image reconstruction and visualisation were carried out using the manufacturer\u0026rsquo;s i-Dixel software (J. Morita). Imaging parameters were standardised between the low-dose and high-dose protocols to ensure geometric consistency and reproducibility across all scans.\u003c/p\u003e \u003cp\u003eThe imaging object was a RANDO anthropomorphic phantom (The Phantom Laboratory, USA), which represented human head and neck anatomy to ensure consistency, reproducibility, and quantitative validity during evaluation\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Two exposure protocols were applied to simulate clinical conditions. The low-dose protocol, used as the noisy input, employed 70 kV and 1 mA, whereas the high-dose protocol, used as the ground-truth reference, employed 90 kV and 10 mA. Both protocols used an identical FOV (140 \u0026times; 100 mm) and scan time (17.5 s) to maintain consistent imaging geometry. Scans were performed in standard imaging mode (CT mode, 360\u0026deg; rotation), with the phantom precisely positioned to ensure spatial alignment between low-dose and high-dose datasets. Each scan yielded 201 axial slices in 16-bit greyscale DICOM format, with an in-plane resolution of 561 \u0026times; 561 pixels.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eEthics and compliance\u003c/h3\u003e\n\u003cp\u003eThis study did not involve human participants. All image data were acquired using a female head RANDO anthropomorphic phantom; therefore, ethical approval and informed consent were not required. AI-assisted language-editing tools were used solely to improve grammatical accuracy and clarity. These tools did not contribute to data analysis, model development, and interpretation of results or conclusions. The authors take full responsibility for the scientific content of the manuscript.\u003c/p\u003e\n\u003ch3\u003eData preprocessing\u003c/h3\u003e\n\u003cp\u003eFrom each CBCT acquisition, 201 paired low-dose and high-dose axial slices were obtained. Ten paired volumes were used for RED-CNN training and validation. In comparison, an additional 10 low-dose volumes that had not been included previously were used for inference or testing to assess generalisation performance. Before processing with the RED-CNN algorithm, DICOM files were imported using the pydicom library and converted to Hounsfield Units (HU) using the RescaleSlope and RescaleIntercept parameters\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Images were adjusted to fit an intensity range of 0\u0026ndash;1 using the HU values. The HU values were limited to the hard-tissue window between \u0026minus;\u0026thinsp;500 and +\u0026thinsp;1500, which is the standard range used in CT imaging for seeing cortical and trabecular structures. This range covers air, soft tissues, and high-density areas in the dentomaxillofacial region\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eEach image was cropped into 512 x 512 pixels to remove air regions and fully cover the RANDO phantom region to reduce computational load and increase training relevance. To improve model generalization, data augmentation was applied through rotation and horizontal/vertical flipping. Augmented images preserved anatomical integrity while producing variations in pixel distributions useful for model learning. To increase the number of training samples, patch extraction was performed, dividing images into overlapping 64x64 pixels using a stride of 8.\u003c/p\u003e \u003cp\u003eThe denoising model was based on the RED-CNN architecture proposed by Chen et al.\u003csup\u003e6\u003c/sup\u003e. The network was implemented using TensorFlow/Keras framework in Python. The input to the network consisted of 2D patches of size 64 x 64 pixels with a single channel (greyscale), corresponding to low-dose dental CBCT images. The RED-CNN architecture comprised five convolutional layers in the encoder path and five corresponding deconvolutional (transposed convolutional) layers in the decoder path. Each convolutional and deconvolutional layer used 96 filters with a kernel size of 5\u0026times;5 and \u0026ldquo;same\u0026rdquo; padding to preserve spatial dimensions. Rectified linear unit (ReLU) activation functions were applied after each convolutional layer. Feature maps from each encoder layer were stored and later added element-wise to the corresponding decoder layer outputs via local residual (skip) connection. A final transposed convolutional layer with a single filter produced the residual, which was then added to the original input patch via a global residual connection.\u003c/p\u003e\n\u003ch3\u003eRED-CNN Architecture\u003c/h3\u003e\n\u003cp\u003eThe model was trained using the Adam optimiser with an initial learning rate of 1 \u0026times; 10\u0026thinsp;\u0026minus;\u0026thinsp;4. A step-decay learning rate schedule was applied, with the learning rate reduced by a factor of 0.991 every 40 epochs\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. The loss function was the MSE between the predicted denoised patches and the corresponding high-dose target patches as expressed in Eq.\u0026nbsp;1.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$Loss=\\frac{1}{N}\\sum_{i}^{N}{‖f\\left({X}_{i}\\right)-{Y}_{i}‖}^{2}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({Y}_{i}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({X}_{i}\\)\u003c/span\u003e\u003c/span\u003e represents high-dose and low-dose CBCT images. The mapping function \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(f\\)\u003c/span\u003e\u003c/span\u003e is used on the low-dose inputs to generate outputs that are as close as possible to the ground truth images \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Training and validation data sets were constructed using 80% or 20% splits of low-dose or high-dose patch pairs, as described in the dataset preparation section. To improve computational efficiency, the training data were organized using the tf.data pipeline, with shuffling and mini-batch training. The batch size was set to 16, and prefetching was enabled to optimise GPU utilization. After training, the final RED-CNN model was used to reconstruct whole low-dose CBCT slices by applying the network to each patch and reassembling the outputs into denoised images.\u003c/p\u003e \u003cp\u003eThe trained RED-CNN model was applied to whole low-dose CBCT slices to produce denoised reconstructions. The model was loaded into TensorFlow/Keras and used to process each slice patch-wise, consistent with the patch extraction strategy used during training. DICOM images were first imported using the pydicom library, converted to Hounsfield Unit (HU) values, and normalized to the range 0\u0026ndash;1 using the same clinical used in the training pipeline.\u003c/p\u003e \u003cp\u003eTo perform reconstruction, each normalized slice was divided into overlapping 64 x 64-pixel patches with a stride of 8. The patches were then input to the RED-CNN model in mini-batches for efficient GPU-accelerated inference. The predicted denoised patches were reassembled into a full-size image using Gaussian-weighted blending to minimise boundary artefacts and ensure smooth patch transitions. After reconstruction, the output image was denormalized back into HU values using the inverse windowing function.\u003c/p\u003e \u003cp\u003eThe reconstructed HU image retained all original DICOM metadata. It was written back into DICOM format by converting HU values into stored pixel values using the original RescaleSlope and RescaleIntercept. This ensured compatibility with CBCT viewing workstations and enabled direct comparison of low-dose, high-dose, and RED-CNN reconstructed images.\u003c/p\u003e \u003cp\u003eFollowing reconstruction, quantitative evaluation of the denoised images was conducted to assess the performance of the RED-CNN model. The test data comprised low-dose CBCT slices excluded from the training phase. The resulting outputs were compared with the high-dose images as ground truth images. This evaluation determined whether the model could reduce noise and produce high-quality CBCT images while maintaining essential anatomical structures.\u003c/p\u003e\n\u003ch3\u003eComparative method: NLM\u003c/h3\u003e\n\u003cp\u003eIn addition, a comparison was made between the denoising results obtained with the RED-CNN algorithm and those obtained with NLM, based on the previously obtained evaluation metrics. NLM is a conventional denoising method that can improve the image quality of low-dose dental CBCT. A comparative evaluation of the performance of RED-CNN and NLM was also conducted, focusing on the computational efficiency of both denoising methods for inference/image reconstruction. To apply NLM to dental CBCT images, we need to determine the optimal input settings. We can determine optimal NLM settings by analyzing the PSNR values. These input settings are found through experiments\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. The input parameters include patch size (5 \u0026times; 5, 7 \u0026times; 7, 9 \u0026times; 9, 11 \u0026times; 11, 15 \u0026times; 15), patch distance (5, 7, 9, 11, 15), and h (0.10, 0.12, 0.14, 0.16, 0.18, 0.20, 0.22, 0.24).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eEvaluation metrics\u003c/h2\u003e \u003cp\u003eThe image quality assessment was based on three primary quantitative metrics, including SDNR, PSNR, dan SSIM\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. SDNR measurement is performed by selecting two different regions of interest (ROI), where ROI\u003csub\u003e1\u003c/sub\u003e is for the object area, and ROI\u003csub\u003e2\u003c/sub\u003e is for the background area, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. This method enabled evaluation of contrast discrimination relative to background noise.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn addition to assessing image quality improvement, this study also examined whether the denoising approach could support the use of lower radiation exposures without compromising diagnostic reliability. The radiation output for the CBCT scan was measured using the DAP, which represents the total X-ray energy delivered across the FOV and serves as a standard measure of patient radiation burden in CBCT examinations. DAP is routinely recorded by most CBCT systems and forms the basis for estimating effective dose across different exposure settings, as reported by Ludlow et.al\u003csup\u003e18\u003c/sup\u003e Comparing the DAP values obtained from the high-dose and low-dose protocols allowed us to evaluate whether diagnostic-quality images could still be achieved under substantially reduced radiation levels. Comparing DAP values from the high-dose and low-dose protocols enabled assessment of whether diagnostic-quality images could be maintained under significantly reduced radiation levels. Dose-reduction efficiency was calculated using:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\%Doseefficiency=\\frac{{DAP}_{high}-{DAP}_{low}}{{DAP}_{high}}\\times100\\%$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({DAP}_{high}\\)\u003c/span\u003e\u003c/span\u003e denotes the radiation output of the high-dose protocol, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({DAP}_{low}\\)\u003c/span\u003e\u003c/span\u003e denotes the output of the low-dose protocol.\u003c/p\u003e \u003cp\u003eComputational performance was also evaluated to determine the practicality of each denoising method in a clinical workflow. Metrics included execution time per slice, GPU/ CPU utilization, and memory consumption, enabling comparison between GPU-accelerated deep learning inference and CPU-based NLM processing.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eThe visual quality of the high-dose CBCT image (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA) was markedly superior to that of the low-dose acquisition (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB), which exhibited increased noise and reduced contrast. This observation is consistent with fundamental radiographic principles, whereby reductions in tube voltage and tube current decrease radiation output, but simultaneously increase quantum noise and degrade image quality\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eQuantitative evaluation demonstrated that the high-dose CBCT images consistently exhibited high SDNR values, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA. The SDNR range for high-dose images was 12.52\u0026ndash;13.34, indicating that the images had good contrast between tissues and minimal noise. In contrast, the low-dose protocol returned SDNR values of 2.33\u0026ndash;2.41, reflecting substantial image degradation. According to the Rose model, an SDNR greater than 5 is generally required for reliable visual detection\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, supporting the need for denoising enhancement in low-dose CBCT imaging. Application of the RED-CNN model to low-dose CBCT slices resulted in a marked improvement in image quality. SDNR values increased to a range of 27.87\u0026ndash;36.44, with a mean SDNR value of 32.87. This increase indicates substantial suppression of background noise and improved signal differentiation. Although SDNR values exceeded those of the high-dose reference images, this increase primarily reflects variance suppression, a known and expected effect of image denoising. Importantly, no loss of anatomical structures was visually observed, indicating that the enhancement remained diagnostically acceptable.\u003c/p\u003e \u003cp\u003eThe PSNR also improved significantly, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB. Low-dose images exhibited PSNR values of 17.88\u0026ndash;18.20 dB, and RED-CNN denoised images reached 30.10\u0026ndash;30.74 dB, corresponding to an improvement of approximately 12 dB. As PSNR values above 30 dB are generally considered indicative of acceptable diagnostic image quality, these findings indicate that RED-CNN-denoised images approximate high-dose image quality to a substantial extent. This result is consistent with findings reported by Chen et al.\u003csup\u003e6\u003c/sup\u003e who observed PSNR improvements of 10 dB to 13 dB in low-dose CT denoising. SSIM analysis provided further evidence of structural preservation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA, low-dose CBCT images exhibited SSIM values of approximately 0.30, whereas RED-CNN-denoised images achieved SSIM values ranging from 0.72 to 0.75 (mean 0.73). The results indicate substantial structural similarity to high-dose (ground truth) images and show that RED-CNN better preserves anatomical contours and tissue interfaces than its conventional counterparts.\u003c/p\u003e \u003cp\u003eFor comparison, the NLM denoising was also applied. Parameter optimisation was performed through empirical testing, and the highest PSNR was achieved with a patch size of 11 \u0026times; 11, a patch distance of 9, and a filtering parameter \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(h\\)\u003c/span\u003e\u003c/span\u003e of 0.14 (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). NLM improves SDNR to 23.13\u0026thinsp;\u0026plusmn;\u0026thinsp;1.78 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA), PSNR to 27.6\u0026thinsp;\u0026plusmn;\u0026thinsp;0.29 dB (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB), and SSIM to 0.67\u0026thinsp;\u0026plusmn;\u0026thinsp;0.0093 (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA). Although NLM effectively reduced random noise, visual inspection revealed subtle blurring of fine anatomical details, consistent with its non-parametric averaging behaviour. The results indicate improved noise suppression and structural preservation in RED-CNN-denoised images compared with NLM, particularly in regions with high anatomical frequency.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe computational performance of both methods is summarised in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. RED-CNN required an average processing time of 14.1 s per function call (42.2 s per slice), whereas NLM required 48.1 s per function call (144.1 s per slice)Radiation dose output for each CBCT protocol was assessed using the dose-area product (DAP). The high-dose protocol yielded a DAP of 4300 mGy. cm\u003csup\u003e2\u003c/sup\u003e, whereas the low-dose protocol yielded a DAP of 238 mGy.cm\u003csup\u003e2\u003c/sup\u003e, corresponding to a dose reduction of approximately 94%. Despite this substantial dose reduction, RED-CNN successfully restored low-dose image quality to levels approaching those of high-dose images, as reflected by PSNR and SSIM values within clinically acceptable ranges. These findings demonstrate that deep learning-based denoising can effectively mitigate image quality degradation associated with dose reductions. The enhanced image quality enabled by RED-CNN contributes to clinically compatible visualisation of dental and maxillofacial structures even at the lowest radiation dosages. From a radiation protection perspective, the results support optimisation strategies consistent with the ALARA principle, enabling diagnostically acceptable CBCT imaging at substantially reduced radiation doses. This is especially helpful for repeated imaging procedures and for radiosensitive patient populations, such as children.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study demonstrates that RED-CNN provides an effective approach for restoring diagnostically acceptable image quality in low-dose dental CBCT imaging. The model consistently improved noise suppression and structural preservation beyond what was achieved with conventional NLM filtering, enabling high-quality visualisation even when the radiation dose was reduced by approximately 94%. These findings have important implications for dose optimisation in dentomaxillofacial imaging.\u003c/p\u003e \u003cp\u003eThe image quality degradation observed in the raw low-dose images is consistent with established radiographic principles. Reductions in tube voltage or tube current decrease photon fluence, leading to increased quantum noise, reduced contrast resolution, and obscuration of fine anatomical detail. These limitations were clearly reflected in the low SDNR, PSNR, and SSIM values of the unprocessed low-dose images. The improvements achieved using RED-CNN demonstrate the capability of deep-learning models to mitigate these physical limitations by learning the relationship between noise characteristics and underlying image structures from paired training data.\u003c/p\u003e \u003cp\u003eIn comparison, NLM provided moderate improvements in objective image quality metrics but introduced blurring of high-frequency anatomical structures, which may reduce diagnostic confidence. By contrast, RED-CNN preserved dental and osseous contours while effectively suppressing stochastic noise, supporting its suitability for clinical interpretation. The improved structural fidelity achieved by RED-CNN reflects the advantage of residual learning in retaining edge detail and spatial gradients that are crucial in dental and maxillofacial evaluation.\u003c/p\u003e \u003cp\u003eThe observed reduction in processing time reflects the contrast between GPU-accelerated feed-forward inference in RED-CNN and the iterative non-local patch search employed by CPU-based NLM. The relatively rapid inference time achieved by RED-CNN suggests potential suitability for integration into time-constrained clinical imaging workflows, where efficient image processing is desirable.\u003c/p\u003e \u003cp\u003eFrom a radiation protection perspective, the most important implication of this study is the demonstrated ability to achieve substantial dose reduction while maintaining diagnostically acceptable image quality. A reduction in radiation dose of approximately 94% directly supports the principle of ALARA and ALADAIP. This capability is especially relevant for paediatric imaging, orthodontic follow-up, implant planning, and other clinical scenarios that require repeated CBCT examinations. The findings suggest that deep learning-based denoising may enable the implementation of ultra-low-dose CBCT protocols without compromising diagnostic utility.\u003c/p\u003e \u003cp\u003eOverall, the results indicate that RED-CNN is a promising tool for enabling significant dose reduction in dental CBCT imaging while ensuring high-quality imaging. Future studies should investigate the generalisability of the model across different CBCT systems, anatomical regions, and clinical conditions, as well as its potential integration into vendor software for real-time reconstructions.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe study demonstrated that the RED-CNN deep learning architecture provides an effective solution for noise suppression and low-dose dental CBCT imaging. The proposed approach substantially improved image quality, with SDNR values exceeding 5 and reaching up to 25, PSNR increasing by more than 10 dB to values above 30 dB, and SSIM rising to approximately 0.7, indicating substantial similarity to high-dose (ground-truth) images. Compared with conventional NLM denoising, RED-CNN consistently outperformed NLM across all quantitative image quality metrics, including SDNR, PSNR, and SSIM. In addition, RED-CNN achieved faster computational performance, demonstrating superior efficiency in both noise suppression and structural preservation.\u003c/p\u003e \u003cp\u003eThese findings highlight the robustness of deep learning-based approaches relative to traditional patch-based denoising methods. Optimisation of the low-dose CBCT acquisition protocol resulted in an approximately 94% reduction in radiation exposure compared with the high-dose protocol, without modification of other acquisition parameters such as FOV or scan time. Together, these results support the feasibility of combining low-dose CBCT protocols with advanced deep learning denoising to maintain diagnostic utility while adhering to the ALARA principle. Overall, RED-CNN represents a promising and clinically relevant approach for enhancing low-dose CBCT imaging while enabling substantial radiation dose reduction without compromising diagnostic image quality.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding acknowledgement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was conducted without financial support from any public, commercial, or not-for-profit funding agency. All resources utilized in this research were provided by the authors\u0026apos; affiliated institutions.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eA.S.S. collected all data and wrote the main manuscript, M.G. and S.S. developed the algorithm, L.E.L. composed the research idea and design. All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe authors thank the Radiology Unit of Universitas Indonesia Hospital for providing access to the CBCT imaging facilities used in this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eNemtoi A, Czink C, Haba D, Gahleitner A (2013) Cone beam CT: a current overview of devices. Dentomaxillofac Radiol 42(8). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1259/DMFR.20120443\u003c/span\u003e\u003cspan address=\"10.1259/DMFR.20120443\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKazimierczak W, Wajer R, Wajer A, Kazimierczak N, Janiszewska-Olszowska J Enhanced Cone-Beam Computed Tomography Imaging through Deep Learning Model Reconstruction: Noise Reduction and Image Quality Optimization in Dental Diagnostics. Published online 2023. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21203/rs.3.rs-3650822/v1\u003c/span\u003e\u003cspan address=\"10.21203/rs.3.rs-3650822/v1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaasalainen T, Ekholm M, Siiskonen T, Kortesniemi M (2021) Dental cone beam CT: An updated review. Physica Medica Associazione Italiana di Fisica Medica 88:193\u0026ndash;217. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ejmp.2021.07.007\u003c/span\u003e\u003cspan address=\"10.1016/j.ejmp.2021.07.007\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCLEMENT CH, HAMADA N Radiological protection in cone beam 11 computed tomography (CBCT) 12. \u003cem\u003eICRP PUBLICATION\u003c/em\u003e. 1(7):8\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePauwels R (2015) Cone beam CT for dental and maxillofacial imaging: dose matters. Radiat Prot Dosimetry 165(1\u0026ndash;4):156\u0026ndash;161\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen H, Zhang Y, Kalra MK et al (2017) Low-Dose CT with a residual encoder-decoder convolutional neural network. IEEE Trans Med Imaging 36(12):2524\u0026ndash;2535. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TMI.2017.2715284\u003c/span\u003e\u003cspan address=\"10.1109/TMI.2017.2715284\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCokrokusumo HB, Hariyati I, Lubis LE, Prajitno P, Soejoko DS (2022) Noise Suppression of Computed Tomography (CT) Images Using Residual Encoder-Decoder Convolutional Neural Network (RED-CNN). Atom Indonesia 48(3):171\u0026ndash;178. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.17146/AIJ.2022.1113\u003c/span\u003e\u003cspan address=\"10.17146/AIJ.2022.1113\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLubis LE, Cokrokusumo HB, Basith RA, Lestariningsih I, Prajitno P, Soejoko DS (2021) Noise reduction of three-dimensional rotational angiography (3DRA) images using residual encoder-decoder convolutional neural network (RED-CNN). In: \u003cem\u003eAIP Conference Proceedings\u003c/em\u003e. Vol 2346. American Institute of Physics Inc.; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1063/5.0048077\u003c/span\u003e\u003cspan address=\"10.1063/5.0048077\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBuades A, Coll B, Morel JM (2011) Non-Local Means Denoising. Image Process Line 1:208\u0026ndash;212. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.5201/ipol.2011.bcm_nlm\u003c/span\u003e\u003cspan address=\"10.5201/ipol.2011.bcm_nlm\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMah E, Ritenour ER, Yao H (2020) A review of dental cone-beam CT dose conversion coefficients. Dentomaxillofacial Radiology British Inst Radiology 50(3). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1259/dmfr.20200225\u003c/span\u003e\u003cspan address=\"10.1259/dmfr.20200225\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTheodorakou C, Walker A, Horner K, Pauwels R, Bogaerts R, Jacobs R (2012) Estimation of paediatric organ and effective doses from dental cone beam CT using anthropomorphic phantoms. Br J Radiol 85(1010):153\u0026ndash;160. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1259/bjr/19389412\u003c/span\u003e\u003cspan address=\"10.1259/bjr/19389412\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eH\u0026ouml;rnlund M, Bernhardsson C (2013) \u003cem\u003eORGAN DOSES DETERMINED USING A RANDO PHANTOM FOR DIFFERENT RADIONUCLIDE DEPOSITIONS AND PHOTON ENERGIES\u003c/em\u003e. Vol 11\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMagill D, Beckmann N, Felice MA, Yoo T, Luo M, Mupparapu M (2018) Investigation of dental cone-beam CT pixel data and a modified method for conversion to hounsfield unit (HU). Dentomaxillofacial Radiol 47(2). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1259/dmfr.20170321\u003c/span\u003e\u003cspan address=\"10.1259/dmfr.20170321\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu L, Vrieze TJ, Bruesewitz MR et al (2010) Dose and image quality evaluation of a dedicated cone-beam CT system for high-contrast neurologic applications. Am J Roentgenol 194(2). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2214/AJR.09.2951\u003c/span\u003e\u003cspan address=\"10.2214/AJR.09.2951\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKingma DP, Ba J, Adam (2014) A Method for Stochastic Optimization. Published online Dec 22, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/1412.6980\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/1412.6980\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiregar S, Nagaoka R, Ul Haq I, Saijo Y (2018) Non local means denoising in photoacoustic imaging. Jpn J Appl Phys 57(7). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7567/JJAP.57.07LB06\u003c/span\u003e\u003cspan address=\"10.7567/JJAP.57.07LB06\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Z, Bovik AC, Sheikh HR, Simoncelli EP (2004) Image quality assessment: From error visibility to structural similarity. IEEE Trans Image Process 13(4):600\u0026ndash;612. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2003.819861\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2003.819861\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLudlow JB, Timothy R, Walker C et al (2015) Ffective dose of dental CBCT - A meta analysis of published data and additional data for nine CBCT units. Dentomaxillofacial Radiology British Inst Radiology 44(1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1259/dmfr.20140197\u003c/span\u003e\u003cspan address=\"10.1259/dmfr.20140197\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLorenzoni DC, Bolognese AM, Garib DG, Guedes FR, Sant\u0026rsquo;Anna EF (2012) Cone-beam computed tomography and radiographs in dentistry: Aspects related to radiation dose. \u003cem\u003eInt J Dent\u003c/em\u003e. Preprint posted online. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1155/2012/813768\u003c/span\u003e\u003cspan address=\"10.1155/2012/813768\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBushberg JT (2012) The Essential Physics of Medical Imaging. Wolters Kluwer Health/Lippincott Williams \u0026amp; Wilkins\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMonnin P, Gnesin S, Verdun FR, Marshall NW (2019) Generalized SDNR analysis based on signal and noise power. Physica Med 64:10\u0026ndash;15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ejmp.2019.06.005\u003c/span\u003e\u003cspan address=\"10.1016/j.ejmp.2019.06.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1. Computational performance between RED-CNN and NLM\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"637\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eComputational Aspect\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRED-CNN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNLMD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eExecution time per slice\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003e\u0026plusmn;42,2 s\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003e\u0026plusmn;144,1 s\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eMean call\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003e\u0026plusmn;14,1 s\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003e\u0026plusmn;48 s\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eRepeatitions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eExecution mode\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003eGPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003eCPU\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eGPU utilization\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003eAverage 8%, \u0026nbsp;peak 14%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003e0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eMemory usage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003e75 MB RAM + 3,9-4,0 VRAM GPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003e6,5 MB RAM + 0,34 VRAM\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eComputational model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003eDeep Learning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003eAlgorithmic, Non-learning\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eComputational efficiency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003eHigh efficiency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003eLower efficiency\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 36.4207%;\"\u003e\n \u003cp\u003eHardware requirements\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 30.2983%;\"\u003e\n \u003cp\u003eGPU recommended for optimal performance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.281%;\"\u003e\n \u003cp\u003eCPU-only is sufficient, but with a prolonged processing time\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"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":"physical-and-engineering-sciences-in-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"apes","sideBox":"Learn more about [Physical and Engineering Sciences in Medicine](http://link.springer.com/journal/13246)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/apes/default.aspx","title":"Physical and Engineering Sciences in Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"CBCT, low-dose imaging, image denoising, RED-CNN, NLM","lastPublishedDoi":"10.21203/rs.3.rs-8930014/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8930014/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eObjective\u003c/h2\u003e \u003cp\u003eTo evaluate the effectiveness of Residual Encoder-Decoder Convolutional Neural Network (RED-CNN) in reducing noise and improving image quality in low-dose dental Cone-Beam Computed Tomography (CBCT), and to compare its performance with the conventional Non-Local Means (NLM) denoising algorithm.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA female head RANDO phantom was scanned using a dental CBCT system with high-dose protocol (90 kV, 10 mA) to obtain ground-truth images and a low-dose protocol (70 kV, 1 mA) to generate noisy datasets. The RED-CNN model was trained and validated on paired low-dose and high-dose images, and tested on unseen data to assess generalization performance. Quantitative evaluation included Signal Difference-to-Noise Ratio (SDNR), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), computational performance, and dose-reduction assessment using the Dose Area Product (DAP).\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eBoth methods reduced noise and enhanced the quality of low-dose dental CBCT images; however, RED-CNN consistently outperformed NLM across all quantitative metrics. RED-CNN achieved SDNR\u0026thinsp;\u0026gt;\u0026thinsp;25, PSNR\u0026thinsp;\u0026gt;\u0026thinsp;30 dB, and SSIM of 0.73, demonstrating improved noise suppression and preservation of anatomical structures, and also provided faster GPU-based inference. Dose analysis showed that the low-dose protocol reduced exposure by 94% while maintaining acceptable diagnostic quality.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThese findings showed that low-dose dental CBCT may achieve clinically acceptable image quality when enhanced using RED-CNN, as the method effectively suppresses noise while preserving essential anatomical detail.\u003c/p\u003e","manuscriptTitle":"Enhancing Image Quality of Low-Dose Dental CBCT Using Residual Encoder- Decoder Convolutional Neural Network (RED-CNN): A Comparative Study with Non-Local Means Denoising","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-19 15:58:42","doi":"10.21203/rs.3.rs-8930014/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-30T08:49:11+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-08T10:12:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"140648161514022387879056169578630602859","date":"2026-03-17T14:51:14+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-16T06:13:48+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-28T11:17:10+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-28T11:15:47+00:00","index":"","fulltext":""},{"type":"submitted","content":"Physical and Engineering Sciences in Medicine","date":"2026-02-21T02:21:31+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"physical-and-engineering-sciences-in-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"apes","sideBox":"Learn more about [Physical and Engineering Sciences in Medicine](http://link.springer.com/journal/13246)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/apes/default.aspx","title":"Physical and Engineering Sciences in Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"84fbd09b-f0e9-46ec-8859-c72903764d9b","owner":[],"postedDate":"March 19th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-04-30T08:49:11+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-04-30T09:03:19+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-19 15:58:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8930014","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8930014","identity":"rs-8930014","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}