3D transformer-based dose prediction in HDR brachytherapy for cervical cancer

3D transformer-based dose prediction in HDR brachytherapy for cervical cancer | 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 3D transformer-based dose prediction in HDR brachytherapy for cervical cancer Weiwei Guo, Wanwei Jian, Lin Zhu, Bailin Zhang, Qiang He, Geng Yang, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5069942/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 20 Jan, 2026 Read the published version in BMC Cancer → Version 1 posted 3 You are reading this latest preprint version Abstract Background The clinical process of high-dose-rate brachytherapy (HDRBT) is time-consuming and reliant on user expertise and preference. Deep learning-based dose prediction can act as a quality assurance (QA) tool to identify suboptimal needle placement and enhance treatment planning efficiency in HDR interstitial brachytherapy. This study aims to introduce a 3D transformer-based deep learning method to automatically predict the dose distribution in HDR brachytherapy for cervical cancer. Methods We retrospectively analyzed 96 CT-based treatment plans from 24 cervical cancer patients who underwent interstitial HDRBT with needle insertion. The transformer mechanism was integrated into a convolutional neural network (CNN) to capture long-distance characteristics and global information. The prediction performance was evaluated by the mean error of Dose-Volume Histogram metrics between clinical and predicted dose maps, gamma analysis, and the Dice similarity coefficient (DSC) of the 1–30 Gy isodose volumes. Results The mean error of D 90 and D 95 for HRCTV was less than 0.1 Gy, and the mean D 2cc for organs at risk (bladder, rectum, and bowel) was under 0.6 Gy. The mean DSC of the 1–30 Gy isodose volume was 0.87. The 3D transformer-based CNN model can predict dose maps that are highly consistent with the clinical treatment plans. Conclusions A novel 3D transformer-based deep learning model was successfully developed for dose prediction in HDR interstitial brachytherapy. This method can automatically generate accurate 3D dose distributions, exhibiting great clinical potential for improving treatment efficiency and standardizing brachytherapy treatment planning. Trial registration: No. ZF2020-084.2; Oct 16, 2020 Brachytherapy dose prediction transformer deep learning cervical cancer Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Background High-dose-rate brachytherapy (HDRBT) along with external beam radiation therapy (EBRT) has been identified as the standard of care for patients with locally advanced cervical cancer [ 1 ]. In clinical practice, the radioactive source utilized in brachytherapy is directly implanted through intracavitary applicators, which leads to a high radiation dose to the tumor and rapid dose falloff at distances from the source [ 2 ]. For complex cases with large volumes or asymmetric sizes of high-risk clinical target volume (HRCTV), brachytherapy is often delivered via freehand interstitial needles to obtain a customized dose distribution [ 3 ]. In such cases, the dose distribution is highly related to patient anatomy and needle geometry, introducing procedural complexity and a time-consuming treatment process [ 4 ]. Additionally, the quality of needle insertion depends on the oncologist’s expertise and preference, which is not standardized and varies among oncologists, potentially affecting the quality of treatment plans. To maintain high-quality treatment planning, there is an urgent need to develop a fast and accurate Quality Assurance (QA) tool to predict the dose distribution for freehand interstitial brachytherapy. Much effort has been devoted to reducing plan variations in HDR brachytherapy. For example, knowledge-based planning (KBP) models have been proven to improve plan quality and predict the dose of new patients on the basis of historical information from treatment plans, such as patient anatomical information, distances from the HRCTV to organs at risk (OAR), and other geometric features [ 5 ]. Despite some progress achieved [ 6 – 8 ], KBP-based methods are highly dependent on the selection of handcrafted features and can only predict the DVH of OARs [ 9 ]. In recent years, deep learning-based methods have achieved significant success in dose prediction for external beam radiation therapy (EBRT), and some have been gradually applied to brachytherapy. Kandalan et al. implemented a 3D convolutional neural network (CNN) to predict the dose distribution for patients with prostate cancer treated with volumetric modulated arc therapy (VMAT) [ 10 ]. Duan et al. developed a 3D asymmetrical ResNeSt dose prediction model for patients with esophageal cancer with various prescription levels [ 11 ]. Zhang et al. employed a neural network to predict the D 2cc /D 90 dosimetric index of each OAR for patients with cervical cancer undergoing needle insertion in HDR brachytherapy [ 12 ]. Ma et al. proposed a 3D deep CNN to predict the dose distribution of tandem and ovoid (T&O) treatment plans for cervical cancer brachytherapy [ 13 ]. Although CNN models have been used and have demonstrated superiority for dose prediction in EBRT and T&O brachytherapy, there are still challenges in predicting the dose distribution for interstitial needle insertion. In particular, the locality of the receptive fields in the convolutional layers may lead to a poor ability to learn long-range spatial dependencies and global context [ 14 ]; thus, CNN models may yield suboptimal dose prediction for relatively small regions in the presence of large interpatient variation [ 15 ]. Consequently, it is challenging for CNN models to predict the dose distribution of freehand needle insertions in HDR brachytherapy, due to the complex geometric relationships between needles, such as crossing or touching. To address this issue, hybrid architectures that combine CNNs and self-attention mechanisms have been proposed to effectively enhance nonlocal modeling capabilities. Li et al. [ 16 ] developed a deep neural network that combined the CNN model and the squeeze-and-excitation attention module to predict the 3D dose distribution in HDRBT, which demonstrated a higher level of agreement with clinical plans than the CNN model. Recognizing the intricate nature of 3D dose prediction in HDRBT and the constraints of existing dose prediction methodologies, we propose an effective deep learning model for fast and accurate 3D dose prediction in HDR brachytherapy. Inspired by the tremendous success of the transformer mechanism [ 17 , 18 ], which is entirely based on the self-attention strategy and has a remarkable ability to capture the global context between sequential data, we developed the transformer mechanism as a feature encoder to overcome the limitations of CNN localized receptive fields. The purpose of this study is to investigate the feasibility of a 3D transformer-based deep learning model for dose prediction in HDR interstitial brachytherapy. The prediction performance is estimated via dose analysis, including the quantitative dose difference of Dose-Volume Histogram (DVH) metrics between clinical and predicted dose maps, 3D gamma analysis, the Dice similarity coefficient (DSC) of the 1–30 Gy isodose volumes, and visualized dose distributions. To the best of our knowledge, this work is the first to employ transformer mechanisms to predict the 3D dose distribution of freehand HDR interstitial brachytherapy via needle insertion. 2. Methods The pipeline of the proposed method for predicting the voxel-wise dose distribution in HDR brachytherapy for cervical cancer patients is shown in Figure.1. The overall workflow includes (1) patient selection, (2) image processing, (3) network training, and (4) dosimetric evaluation and analysis. Each step is described in detail in sections 2.1 to 2.4. 2.1 Clinical data and treatment planning This study retrospectively analyzed 96 CT-based treatment plans from 24 patients with different-stages of cervical cancer who were treated with freehand interstitial HDR brachytherapy via needle insertion. The patients’ basic characteristics are presented in Table 1 . The average volume of the HRCTV was 94.6 cm 3 at the time of brachytherapy. Each patient was implanted with 4‒6 trocar stainless steel needles (Elekta AB, diameter of 1.5 mm, length of 200 mm) for a total of 4‒5 fractions of 192 Ir HDR brachytherapy, and delivered with a dose of 6 Gy/fraction on the Flexitron HDR treatment afterloader (Elekta AB, Stockholm, Sweden). During brachytherapy, the radiation oncologist inserted trocar needles into the patient on the basis of the tumor topography, size, and OARs proximity. Subsequently, a CT scan was performed for each patient using a Siemens CT simulator (SOMATOM Sensation Open, Siemens Medical System, Germany) for contour delineation. All the CT images were scanned with a tube voltage of 120 kVp, a tube current of 150 mAs, a reconstructed size of 512×512, and CT acquisition resolutions of 0.6–1.2 mm in-plane and 5 mm longitudinal. The clinical target and OARs, including the high-risk clinical target volume (HRCTV), bladder, rectum, and bowel, were delineated by experienced radiation oncologists based on ESTRO recommendations [ 19 ]. The CT images and contoured structures were then transferred to the Oncentra Brachy treatment planning system (Elekta-Brachy, Veenendaal, Netherlands) for brachytherapy treatment planning. The interstitial needles were manually digitized by experienced medical physicists to determine the trajectory of the radioactive source. The dose calculation algorithm was based on Task Group Report No.43, as recommended by the American Association of Physicists in Medicine (AAPM) [ 20 ]. The treatment planning for all patients was optimized via the Hybrid Inverse treatment Planning and Optimization algorithm (HIPO) [ 21 ], and subsequently manually fine-tuned using graphical optimization tool. Finally, the clinical plan was approved by experienced radiation oncologists. Table 1 Patient characteristics Characteristics Patients Age (years) 53.6 ± 11.0 Volume of HR-CTV (cm 3 ) 94.6 ± 54.3 Prescription dose (Gy) per fraction 6 Gy \(\:\times\:\) 4 ~ 5 fx Number of needles (mean ± std) 4.0 ± 0.6 Clinical stage (FIGO*) (n = 24) ⅠA 6 ⅡA 1 ⅡB 5 ⅢA 2 ⅢB 3 ⅢC 7 Total patients 24 Total plans 96 * FIGO (International Federation of Gynecology and Obstetrics) 2.2 Image processing The overall treatment plan (N = 96) was randomly divided into a training set (N = 81) and a testing set (N = 15). In this study, five structures, including the HRCTV, bladder, rectum, bowel, and needle trajectory, were converted into binary masks and combined as 3D input data. A zero-padding strategy was applied to the z-dimension of each input channel to maintain the same depth of 128. The corresponding dose distribution map with a grid size of 1 mm was considered the ground truth and aligned to the pixel spacing of the binary volumetric structure via trilinear interpolation. In addition, all the input images were cropped to 256×256 pixels at the geometric center of the binary mask to eliminate unnecessary background pixels and address the GPU memory limitations. No data augmentation strategy was utilized because the high-dose distribution may follow the specific spatial relationship of the OARs and strict physical constraints. Finally, a volume size of 256×256×128×5 was fed into the deep learning-based model for HDR brachytherapy dose prediction. 2.3 Transformer-based CNN architecture Inspired by the UNETR model [ 15 ], we reformulated the 3D HDR brachytherapy dose prediction task as a 1D sequence-to-sequence prediction problem. The network architecture of the proposed transformer-based CNN model is depicted in Fig. 2 . We integrated a stack of transformer blocks into a 3D Unet-shaped deep learning network as a feature encoder to learn complex contextual information and long-range dependencies across 3D volumetric images. Subsequently, the feature maps from multiple resolutions of the transformer encoder were merged with the fine-grained feature maps in the upsampling CNN decoder via a skip-connection. Specifically, we first performed a patch partition layer to create a sequence of 3D tokens, and the flattened tokens were projected into a K-dimensional embedding space via a linear projection layer. To retain the spatial context of the extracted patch, 1D learnable positional embedding was added to the projected patch embedding. Thus, the input sequence Z 0 for the transformer encoder can be formulated as: $$\:\begin{array}{c}{Z}_{0}=\left[{x}_{p}^{1}E;{x}_{p}^{2}E;\cdots\:;{x}_{p}^{N}E\right]+{E}_{pos}\#\left(1\right)\end{array}$$ where x p ∈ \(\:{\mathbb{R}}^{N\times\:({P}^{3}\bullet\:C)}\) with each patch size of (p, p, p) and C input channels. N=(H×W×D)/P 3 represented the length of the sequence and (H, W, D) represented the resolution of the 3D input volume. E∈ \(\:{\mathbb{R}}^{({P}^{3}\bullet\:C)\times\:K}\) represented the projected patch embedding and E pos ∈ \(\:{\mathbb{R}}^{N\times\:K}\) represented the 1D position embedding. In this study, the 3D input volume (H×W×D×C)=(256×256×128×5), the patch resolution P = 16, and the embedding size K = 768. As the standard architecture in NLP, each transformer block consists of two main sublayers that follow layer normalization (LN): multihead self-attention (MSA) [ 22 ] and multilayer perceptron (MLP). A residual connection was added to each sublayer to enhance information flow. The encoder consists of several stacked transformer blocks, and the output Z i of the ith block in the transformer encoder can be described as follows: $$\:\begin{array}{c}{Z}_{i}^{{\prime\:}}=MSA\left(\text{L}\text{N}\left({Z}_{i-1}\right)\right)+{Z}_{i-1}\#\left(2\right)\end{array}$$ $$\:\begin{array}{c}{Z}_{i}=MLP\left(\text{L}\text{N}\left({Z}_{i}^{{\prime\:}}\right)\right)+{Z}_{i}^{{\prime\:}}\#\left(3\right)\end{array}$$ where \(\:{Z}_{i}^{{\prime\:}}\) was the output of the MSA layer. i ∈ {1, 2, ……, L} and L = 12 was the total number of transformer blocks in the encoder. For the CNN decoder, the intermediate and bottleneck sequence representations Z i (i ∈{3,6,9,12}) from the transformer encoder were first reshaped to the input space ( \(\:\frac{H}{p}\times\:\frac{W}{p}\times\:\frac{D}{p}\times\:K\) ), and the resolutions were increased by deconvolutional layers with a factor of 2. Furthermore, the transformed feature maps at multiresolutions were combined with a deconvolution layer followed by a 3×3×3 convolutional layer, batch normalization (BN), and rectified linear unit (ReLU) activation functions to adjust the latent dimension. The processed feature maps were then connected to the CNN-based decoder via skip connections. The concatenated feature maps were fed into a decoder block that contained two 3×3×3 convolutional layers, followed by a BN and ReLU, and a deconvolutional layer with a stride of 2 for upsampling to the input resolution. Notably, the decoder employed 6 additional convolutional layers to smoothly decrease the number of filters from 64 to 1 for better preservation of the extracted features [ 13 ], followed by a final 1×1×1 convolution layer to generate the predicted dose map. 2.4 Loss function and implementation details Considering the high dose with steep gradients of dose distribution in HDR brachytherapy [ 23 ], we employed a voxelwise mean square error (MSE) as the loss function. For precise dose prediction, we also calculated the MSE loss in each ROI, including the body, HRCTV, bladder, rectum, bowel, and needle trajectories. The entire loss function of our proposed transformer-based CNN is expressed as: $$\:\begin{array}{c}{L}_{ROI}=\frac{1}{n}\sum\:_{i=1}^{n}({\text{Y}}_{i}-{\widehat{Y}}_{i}{)}^{2}\#\left(4\right)\end{array}$$ $$\:\begin{array}{c}Loss={\lambda\:}_{1}{L}_{body}+{\lambda\:}_{2}{L}_{HRCTV}+{\lambda\:}_{3}{L}_{bladder}+{\lambda\:}_{4}{L}_{rectum}+{\lambda\:}_{5}{L}_{bowel}+{\lambda\:}_{6}{L}_{needles}\#\left(5\right)\end{array}$$ where \(\:{\text{Y}}_{i}\) denoted the ground truth dose value, \(\:{\widehat{Y}}_{i}\) denoted the predicted dose value, and n denoted the number of ROI points. λ i denoted the hyperparameters that balanced the five ROIs. This weighting factor allowed the network to place greater emphasis on learning the challenging regions of the dose prediction task. The coefficients λ1 = λ2 = 3, and λ3 = λ4 = λ5 = λ6 = 1 were considered. The proposed network was implemented using the PyTorch framework on an Intel Xeon Bronze 3206R CPU @ 1.90 GHz, NVIDIA Quadro RTX 6000 with 24 GB of memory. In addition, we employed the adaptive moment estimation (Adam) optimizer with a learning rate of 1×10 − 3 to minimize the MSE loss function. We trained the neural network for 500 epochs with an early stopping strategy, and the batch size was set to 1. 2.5 Dosimetric evaluation and statistical analysis For dosimetric analysis, the mean error (ME) of the dosimetric parameters between the clinical and predicted dose maps was used to assess the prediction accuracy (ME = clinical-prediction). We compared the DVH parameters D 90 and D 95 for the HRCTV, D 1cc , and D 2cc for the OARs. The dose volume reporting parameters of the ESTRO recommendations [ 19 ] (V 100 , V 150 , and V 200 of the HRCTV) were used to calculate the dose distribution metrics to characterize the dose coverage, including the coverage index (CI = V 100 /V pres ), homogeneity index (HI = 1-V 150 /V 100 ), overdose volume index (ODI = V 200 /V 100 ) and dose nonuniformity ratio (DNR = V 150 /V 100 ) [ 24 ]. D x (Gy) represents the dose received by x% of the HRCTV. D xcc represents the dose to the most exposed x cc volume of the OARs. V x represents the percentage volume of the corresponding ROI that receives at least x% of the prescribed dose. V pres represents the volume of the prescribed dose region. In addition, a global 3D gamma analysis under the 3%/3 mm criterion was used to further evaluate the prediction accuracy. We calculated the Dice similarity coefficient (DSC) from 1 to 30 Gy to evaluate the isodose volume similarity between the transformer-based CNN model prediction and the clinical ground truth. The spatial dose distributions and DVH comparisons were visualized for qualitative evaluation. Furthermore, statistical analysis using paired t-tests was performed to assess the differences in dosimetric parameters between the network prediction and the clinical plan. A p-value less than 0.05 was considered statistically significant. 3. Results The transformer-based CNN network took approximately 14 h to train a total of 400 epochs, and 8.5 s to predict a 3D dose distribution for cervical cancer brachytherapy. Table 2 tabulated the mean and standard deviation of the dose differences for all the test sets. The 3D transformer-based CNN network achieved good dose agreement with the clinical treatment plans, with average prediction errors of D 90 = 0.08 ± 0.32 Gy (p = 0.41) and D 95 =-0.03 ± 0.34 Gy (p = 0.78) for the HRCTV. The mean errors of D 1cc and D 2cc for all OARs (bladder, rectum, and bowel) were less than 0.6 Gy. In particular, the mean errors of D 1cc and D 2cc for the bowel were less than 0.2 Gy. Moreover, the 3D gamma evaluation under the 3%/3 mm criterion was 92.73 ± 3.82%, with the following dose distribution metrics: CI=-0.02 ± 0.05, HI=-0.04 ± 0.08, ODI = 0.04 ± 0.11 and DNR = 0.04 ± 0.08. Overall, there was no statistically significant difference in these dosimetric metrics between the deep learning predicted dose map and the clinical ground truth (p > 0.05), except for D 1cc and D 2cc of the rectum. Table 2 showed that the dose coverage of the HRCTV and the dose falling outside the target volume in the predicted dose map are highly consistent with clinical dose map. Figure 3 (a) and (b) showed an axial view of the clinical and predicted dose maps for the two test cases, with the corresponding DVHs and isodose lines depicted. The DVHs of the predicted dose maps matched well with the clinical ground truth. Additionally, the violin plot of different DVH metrics was presented in Fig. 3 (c) to find out the distribution of dose differences for each metric. Figure 4 showed the DSC calculation for evaluating the isodose volume similarity (1–30 Gy) among all the testing cases. The average DSC of all testing cases for the 1–30 Gy isodose level was 0.87 with a range of 0.77–0.94, indicating high dose conformity between the predicted dose distribution and that in clinical practice. Table 2 Dose differences (mean ± SD) of different dosimetric parameters of all the test cases. Dosimetric index 3D transformer-based CNN models p-value HRCTV D 90 (Gy) 0.08 ± 0.32 0.41 D 95 (Gy) -0.03 ± 0.34 0.78 Bladder D 1cc 0.6 ± 1.21 0.27 D 2cc 0.54 ± 0.89 0.16 Rectum D 1cc 0.58 ± 0.49 0.01 D 2cc 0.57 ± 0.42 0.01 Bowel D 1cc 0.16 ± 1.03 0.77 D 2cc 0.19 ± 0.93 0.69 Dose distribution metrics CI -0.02 ± 0.05 0.17 HI -0.04 ± 0.08 0.09 ODI 0.04 ± 0.11 0.24 DNR 0.04 ± 0.08 0.09 Gamma pass rate 3%/3mm 92.73 ± 3.82 / 4. Discussion In this study, we proposed a 3D transformer-based deep learning framework to predict accurate dose distributions for cervical cancer HDR brachytherapy. Taking advantage of the hybrid CNN-transformer architecture, our proposed network not only encodes a strong global context and learns long-range dependencies via the transformer mechanism, but also effectively leverages high-resolution spatial details via CNNs. Furthermore, we developed an MSE loss function that supervised the dose accuracy within body and region of interest (ROI) to improve network performance. Our proposed network generated a real-time dose map based on patient anatomy and needle geometry, which has the clinical potential to help oncologists adjust interstitial implants to avoid an insufficient dose to the target volume and provide better protection of the OARs. This approach can serve as an automatic QA tool in clinical practice to reduce human error during the operation process and further improve treatment efficacy. Dose prediction in interstitial brachytherapy is a more complex process than in EBRT, as its dose distribution is highly constrained by patient anatomy and insertion geometry. Highly inhomogeneous dose distributions and steep dose gradients within the HRCTV are great challenges that hamper the development of automatic dose prediction for HDRBT. Although few studies have investigated deep learning-based methods to facilitate 3D dose prediction for HDR brachytherapy, they have focused mostly on tandem and ovoid (T&O) or tandem-and-ring (T&R) applicators, with little attention given to free-hand interstitial needles. Previously, Cortes et al. developed a 3D CNN to predict dose maps for T&O brachytherapy and reported a mean dose difference of -0.09 ± 0.67 Gy for HRCTV D 90 , -0.17 ± 0.67 Gy for bladder D 2cc , and − 0.04 ± 0.46 Gy for rectum D 2cc [ 6 ]. Ma et al. employed a 3D CNN to predict 3D dose distributions for T&O brachytherapy and proposed a dose difference of 1% for CTV V 100 , 0.2 Gy for bladder D 2cc , 0.25 Gy for rectum D 2cc and 0.11 Gy for bowel D 2cc [ 13 ]. Li et al. demonstrated an attention-based deep learning method for 3D brachytherapy dose prediction and predicted average dose differences of 0.37 ± 0.25 Gy for HRCTV D 90 , 0.23 ± 0.14 Gy for bladder D 2cc , and 0.28 ± 0.20 Gy for rectum D 2cc [ 16 ]. Obviously, dose prediction for interstitial brachytherapy via needle insertion is a more challenging undertaking than for T&O brachytherapy. This is attributable to the increased complexity of the needle geometry, and the varying number of needles may further complicate the dose distribution within the HRCTV. Comparatively, our proposed 3D transformer-based methods presented mean dose differences of 0.08 ± 0.32 Gy for HRCTV D 90 , 0.54 ± 0.89 Gy for bladder D 2cc , 0.57 ± 0.42 Gy for rectum D 2cc, and 0.19 ± 0.93 Gy for bowel D 2cc (Table 2 ). Besides, to further evaluate the dose accuracy, gamma analysis was performed and achieved a 3%/3 mm gamma passing rate of 92.73 ± 3.82% between the network prediction and the clinical ground truth. As shown in Fig. 3 (a), visualization of the 1–30 Gy isodose level revealed that the predicted dose map inherits a similar dose distribution to the clinical dose map. High average DSC (> 0.75) between the predicted dose distribution and the clinical ground truth was observed at the 1–30 Gy isodose level (Fig. 4 ). Our proposed transformer-based CNN method can predict 3D dose distributions consistent with clinical treatment planning. There are several limitations in our study. First, the entire transformer-based network is designed for 3D volume to better preserve the specific spatial relationships between slices. However, this resulted in a notable increase in the computational intensity of the GPU. To address this, network hyperparameters such as batch size, the number of transformer blocks, and the channel of convolutional layers were adjusted to reduce the computational complexity, which may lead to suboptimal prediction in certain cases. Second, the proposed network still requires patient contours as input and may present potential uncertainty in manual delineation. Third, in accordance with the findings of Kallis et al. [ 25 ], a larger dose difference (isodose DSC of 0.77) was observed at a small dose level (1 Gy), which may be due to the steep dose gradients caused by the single dwell positions. Consequently, more clinical plan data need to be employed to achieve better training of the deep learning model. Moreover, considering that only one institution was included in this project and the varying treatment protocols of different institutions, multi-institutional involvement is essential for the development of a robust and accurate dose prediction model. This study is tailored specifically to fully interstitial situations. In future work, the proposed approach can be expanded to facilitate hybrid applicator scenarios to develop clinical benefits. The dwell time and dwell position of the radioactive source based on the predicted dose distribution can be further calculated to improve the automation of the HDR brachytherapy treatment process. 5. Conclusion In this work, we proposed and evaluated a 3D transformer-based CNN method that can serve as a QA tool for needle insertion in HDR interstitial brachytherapy. Dosimetric evaluation showed that the predicted dose distributions were in good agreement with the clinical dose plans. This method has the clinical potential to facilitate brachytherapy treatment workflow and standardize brachytherapy treatment planning. Abbreviations HDRBT High-dose-rate brachytherapy EBRT External beam radiation therapy HRCTV High risk clinical target volume OARs Organs at risk T&O Tandem and ovoid T&R Tandem-and-ring FIGO International Federation of Gynecology and Obstetrics KBP knowledge-based planning CNN Convolutional neural network HIPO Hybrid Inverse treatment Planning and Optimization algorithm QA Quality assurance DSC Dice similarity coefficient DVH Dose-Volume Histogram CI Coverage index HI Homogeneity index ODI Overdose volume index DNR Dose nonuniformity ratio Declarations Ethics approval and consent to participate This study was reviewed and approved by the Institutional Review Board of the Second Affiliated Hospital of Guangzhou University of Chinese Medicine (No. ZF2020-084.2). It was conducted in accordance with the Declaration of Helsinki 1964 and its later amendments or comparable ethical standards. Consent was waived due to the retrospective nature of this study. Consent for publication Not applicable Competing interests The authors declare that they have no competing interests Funding This study is supported by grants from the National Natural Science Foundation of China (Grant No. 82172020), the State Key Laboratory of Traditional Chinese Medicine Syndrome (Grant No. QZ2023ZZ05), and the Ministry of Education Industry-Academic Cooperation Project (Grant No. 230806071142911). Author Contribution Weiwei Guo and Wanwei Jian performed conceptualization, methodology, and writing of the original draft. Lin Zhu and Bailin Zhang contributed data curation and validation. Qiang He performed validation, reviewing and editing of the manuscript. Geng Yang supervised the study. Xuetao Wang contributed to the conceptualization and project administration. All the authors have read and approved the final manuscript. Weiwei Guo and Wanwei Jian contributed equally as the first authors. Acknowledgements Not applicable Data Availability The datasets used during the current study are available from the corresponding author on reasonable request. References Sturdza A, Pötter R, Fokdal LU, Haie-Meder C, Tan LT, Mazeron R, et al. Image guided brachytherapy in locally advanced cervical cancer: Improved pelvic control and survival in RetroEMBRACE, a multicenter cohort study. Radiotherapy Oncology: J Eur Soc Therapeutic Radiol Oncol. 2016;120:428–33. https://doi.org/10.1016/j.radonc.2016.03.011 . Chargari C, Deutsch E, Blanchard P, Gouy S, Martelli H, Guérin F, et al. Brachytherapy: An overview for clinicians. Cancer J Clin. 2019;69:386–401. https://doi.org/10.3322/caac.21578 . Nomden CN, de Leeuw AAC, Moerland MA, Roesink JM, Tersteeg RJHA, Jürgenliemk-Schulz IM. Clinical use of the Utrecht applicator for combined intracavitary/interstitial brachytherapy treatment in locally advanced cervical cancer. Int J Radiat Oncol Biol Phys. 2012;82:1424–30. https://doi.org/10.1016/j.ijrobp.2011.04.044 . Nomden CN, de Leeuw AAC, Moerland MA, Roesink JM, Tersteeg RJHA, Jürgenliemk-Schulz IM. Clinical use of the Utrecht applicator for combined intracavitary/interstitial brachytherapy treatment in locally advanced cervical cancer. Int J Radiat Oncol Biol Phys. 2012;82:1424–30. https://doi.org/10.1016/j.ijrobp.2011.04.044 . Ge Y, Wu QJ. Knowledge-based planning for intensity-modulated radiation therapy: A review of data-driven approaches. Med Phys. 2019;46:2760–75. https://doi.org/10.1002/mp.13526 . Cortes KG, Kallis K, Simon A, Mayadev J, Meyers SM, Moore KL. Knowledge-based three-dimensional dose prediction for tandem-and-ovoid brachytherapy. Brachytherapy. 2022;21:532–42. https://doi.org/10.1016/j.brachy.2022.03.002 . Kallis K, Mayadev J, Kisling K, Brown D, Scanderbeg D, Ray X, et al. Knowledge-based dose prediction models to inform gynecologic brachytherapy needle supplementation for locally advanced cervical cancer. Brachytherapy. 2021;20:1187–99. https://doi.org/10.1016/j.brachy.2021.07.001 . Yusufaly TI, Kallis K, Simon A, Mayadev J, Yashar CM, Einck JP, et al. A knowledge-based organ dose prediction tool for brachytherapy treatment planning of patients with cervical cancer. Brachytherapy. 2020;19:624–34. https://doi.org/10.1016/j.brachy.2020.04.008 . Wen L, Xiao J, Zeng J, Zu C, Wu X, Zhou J, et al. Multi-level progressive transfer learning for cervical cancer dose prediction. Pattern Recogn. 2023;141:109606. https://doi.org/https://doi.org/10.1016/j.patcog.2023.109606 . Kandalan RN, Nguyen D, Rezaeian NH, Barragán-Montero AM, Breedveld S, Namuduri K, et al. Dose prediction with deep learning for prostate cancer radiation therapy: Model adaptation to different treatment planning practices. Radiotherapy Oncology: J Eur Soc Therapeutic Radiol Oncol. 2020;153:228–35. https://doi.org/10.1016/j.radonc.2020.10.027 . Duan Y, Wang J, Wu P, Shao Y, Chen H, Wang H, et al. AS-NeSt: A Novel 3D Deep Learning Model for Radiation Therapy Dose Distribution Prediction in Esophageal Cancer Treatment With Multiple Prescriptions. Int J Radiation Oncology*Biology*Physics. 2024;119:978–89. https://doi.org/https://doi.org/10.1016/j.ijrobp.2023.12.001 . Zhang H-W, Zhong X-M, Zhang Z-H, Pang H-W. Dose prediction of organs at risk in patients with cervical cancer receiving brachytherapy using needle insertion based on a neural network method. BMC Cancer. 2023;23:385. https://doi.org/10.1186/s12885-023-10875-6 . Ma M, Kidd E, Fahimian BP, Han B, Niedermayr TR, Hristov D, et al. Dose Prediction for Cervical Cancer Brachytherapy Using 3-D Deep Convolutional Neural Network. IEEE Trans Radiation Plasma Med Sci. 2022;6:214–21. https://doi.org/10.1109/TRPMS.2021.3098507 . Chen J, Lu Y, Yu Q, Luo X, Adeli E, Wang Y et al. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation. ArXiv. 2021; abs/2102.04306. Hatamizadeh A, Yang D, Roth HR, Xu D. UNETR: Transformers for 3D Medical Image Segmentation. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2021:1748–58. Li Z, Yang Z, Lu J, Zhu Q, Wang Y, Zhao M, et al. Deep learning-based dose map prediction for high-dose-rate brachytherapy. Phys Med Biol. 2023;68. https://doi.org/10.1088/1361-6560/acecd2 . Vaswani A, Shazeer NM, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is All you Need. Neural Information Processing Systems; 2017. Lu W, Jiang J, Tian H, Gu J, Lu Y, Yang W, et al. Asymmetric convolution Swin transformer for medical image super-resolution. Alexandria Eng J. 2023;85:177–84. https://doi.org/https://doi.org/10.1016/j.aej.2023.11.044 . Pötter R, Haie-Meder C, Van Limbergen E, Barillot I, De Brabandere M, Dimopoulos J, et al. Recommendations from gynaecological (GYN) GEC ESTRO working group (II): concepts and terms in 3D image-based treatment planning in cervix cancer brachytherapy-3D dose volume parameters and aspects of 3D image-based anatomy, radiation physics, radiobiolog. Radiotherapy and Oncology. J Eur Soc Therapeutic Radiol Oncol. 2006;78:67–77. https://doi.org/10.1016/j.radonc.2005.11.014 . Rivard MJ, Coursey BM, DeWerd LA, Hanson WF, Huq MS, Ibbott GS, et al. Update of AAPM Task Group 43 Report: A revised AAPM protocol for brachytherapy dose calculations. Med Phys. 2004;31:633–74. https://doi.org/10.1118/1.1646040 . Karabis A, Giannouli S, Baltas D. 40 HIPO: A hybrid inverse treatment planning optimization algorithm in HDR brachytherapy. Radiother Oncol 2005;76. Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ArXiv 2020;abs/2010.1. Tanderup K, Ménard C, Polgar C, Lindegaard JC, Kirisits C, Pötter R. Advancements in brachytherapy. Adv Drug Deliv Rev. 2017;109:15–25. https://doi.org/https://doi.org/10.1016/j.addr.2016.09.002 . Kaur G, Garg P, Srivastava AK, Gaur G, Sheetal, Grover R, et al. Dosimetric and radiobiological evaluation of treatment plan for cervical cancer high-dose-rate intracavitary brachytherapy. J Contemp Brachytherapy. 2022;14:253–9. https://doi.org/10.5114/jcb.2022.117729 . Kallis K, Moore LC, Cortes KG, Brown D, Mayadev J, Moore KL et al. Automated treatment planning framework for brachytherapy of cervical cancer using 3D dose predictions. Phys Med Biol 2023;68. https://doi.org/10.1088/1361-6560/acc37c Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 20 Jan, 2026 Read the published version in BMC Cancer → Version 1 posted Editor assigned by journal 12 Sep, 2024 Submission checks completed at journal 12 Sep, 2024 First submitted to journal 11 Sep, 2024 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-5069942","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":353258264,"identity":"85bbb458-4174-45fc-beb7-00d7637a2bbf","order_by":0,"name":"Weiwei Guo","email":"","orcid":"","institution":"Department of Radiation Therapy, The Second Affiliated Hospital, Guangzhou University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Weiwei","middleName":"","lastName":"Guo","suffix":""},{"id":353258265,"identity":"f2825376-b0bf-4c41-8235-1809784cdf24","order_by":1,"name":"Wanwei Jian","email":"","orcid":"","institution":"Department of Radiation Therapy, The Second Affiliated Hospital, Guangzhou University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Wanwei","middleName":"","lastName":"Jian","suffix":""},{"id":353258266,"identity":"2cf34990-2572-479c-9edf-ee203d22396c","order_by":2,"name":"Lin Zhu","email":"","orcid":"","institution":"Department of Radiation Therapy, The Second Affiliated Hospital, Guangzhou University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Zhu","suffix":""},{"id":353258267,"identity":"dd674413-5551-4942-8685-c8f6a9962d67","order_by":3,"name":"Bailin Zhang","email":"","orcid":"","institution":"Department of Radiation Therapy, The Second Affiliated Hospital, Guangzhou University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Bailin","middleName":"","lastName":"Zhang","suffix":""},{"id":353258268,"identity":"459325e9-377b-4246-a09a-59fad4f74612","order_by":4,"name":"Qiang He","email":"","orcid":"","institution":"Department of Radiation Therapy, The Second Affiliated Hospital, Guangzhou University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Qiang","middleName":"","lastName":"He","suffix":""},{"id":353258269,"identity":"07d66e1c-1ad6-44e6-bd82-185d9ff2fa6f","order_by":5,"name":"Geng Yang","email":"","orcid":"","institution":"Department of Radiation Therapy, The Second Affiliated Hospital, Guangzhou University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Geng","middleName":"","lastName":"Yang","suffix":""},{"id":353258270,"identity":"41652695-41e5-4874-96f6-94edb18bfe50","order_by":6,"name":"Xuetao Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAElEQVRIie3PsUoDMRzH8YSDuOS4wSUHRV8hpRARjp6PohQyBUdxkhyFdCm4evgULuKYI8MtsV0DXQ4c7XDBtaA3C6YdheYz/Ifw+w4BIIr+N1h1flfg7EQeticAJGZcSz7Kl/rgBPHTVJqCuqvwlLbv5jNVxUO2WDKQvq0xcAD2XgQSe8sva8UJsZZ1ud1g+CyTvH79O2FaMOqVIcCJC0rRBicjjZI0lKy3jN6ob3LuBCPXaIXRcMOJE5POK02o45w0SmO8LyndlsGn1Sx/scaMpZ1hgpt58C/5o5h84btpdtZW1cfuflqW7bzpfSAZIPLrAcrgfpD0+xZRFEVH7gfYElaeGKsmHgAAAABJRU5ErkJggg==","orcid":"","institution":"State Key Laboratory of Traditional Chinese Medicine Syndrome, The Second Affiliated Hospital of Guangzhou University of Chinese Medicine","correspondingAuthor":true,"prefix":"","firstName":"Xuetao","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2024-09-11 09:33:07","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5069942/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5069942/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12885-026-15592-4","type":"published","date":"2026-01-20T15:57:22+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":70947312,"identity":"c5be3d20-e4a6-4679-bd42-75f782b74e98","added_by":"auto","created_at":"2024-12-09 13:09:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":712774,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic flowchart of dose prediction for HDR brachytherapy using the 3D transformer-based CNN method.\u003c/p\u003e","description":"","filename":"Figure1.flowchart.png","url":"https://assets-eu.researchsquare.com/files/rs-5069942/v1/d8e60fa77e847594f04a1e6e.png"},{"id":70945873,"identity":"30d0e1f3-de90-4082-9be6-cf0c61f7742e","added_by":"auto","created_at":"2024-12-09 13:01:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":375393,"visible":true,"origin":"","legend":"\u003cp\u003eThe network architecture of 3D transformer-based CNN model.\u003c/p\u003e","description":"","filename":"Figure2.network.png","url":"https://assets-eu.researchsquare.com/files/rs-5069942/v1/cb7d304f87347d568ff7c1e4.png"},{"id":70947313,"identity":"57f70b49-0b3d-49fb-92bf-80382230e943","added_by":"auto","created_at":"2024-12-09 13:09:14","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":987493,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Dose distribution, isodose lines, and corresponding DVHs of the two test cases. In the DVH graph, the dashed lines represent the DVH derived from the clinical dose distribution, and the dotted lines represent the DVH of the network prediction. (b) Violin plot of dose differences for all testing cases.\u003c/p\u003e","description":"","filename":"Figure3.dose.png","url":"https://assets-eu.researchsquare.com/files/rs-5069942/v1/bc627edd640e186813a15bbd.png"},{"id":70945875,"identity":"613afff7-bceb-4e90-b57a-ad0b75a16569","added_by":"auto","created_at":"2024-12-09 13:01:14","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":127018,"visible":true,"origin":"","legend":"\u003cp\u003eDice similarity coefficient analysis. The shadow indicates the standard deviation. The prescription dose was 6 Gy.\u003c/p\u003e","description":"","filename":"Figure4.IsodoseDSC.png","url":"https://assets-eu.researchsquare.com/files/rs-5069942/v1/c36dec210666a6a7aa4c7019.png"},{"id":101151825,"identity":"66ab4b45-ae97-4592-b151-6536109b7d15","added_by":"auto","created_at":"2026-01-26 16:06:18","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2883687,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5069942/v1/0086d845-13f4-4be3-959e-b2338b651069.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"3D transformer-based dose prediction in HDR brachytherapy for cervical cancer","fulltext":[{"header":"1. Background","content":"\u003cp\u003eHigh-dose-rate brachytherapy (HDRBT) along with external beam radiation therapy (EBRT) has been identified as the standard of care for patients with locally advanced cervical cancer [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. In clinical practice, the radioactive source utilized in brachytherapy is directly implanted through intracavitary applicators, which leads to a high radiation dose to the tumor and rapid dose falloff at distances from the source [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. For complex cases with large volumes or asymmetric sizes of high-risk clinical target volume (HRCTV), brachytherapy is often delivered via freehand interstitial needles to obtain a customized dose distribution [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. In such cases, the dose distribution is highly related to patient anatomy and needle geometry, introducing procedural complexity and a time-consuming treatment process [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Additionally, the quality of needle insertion depends on the oncologist\u0026rsquo;s expertise and preference, which is not standardized and varies among oncologists, potentially affecting the quality of treatment plans.\u003c/p\u003e \u003cp\u003eTo maintain high-quality treatment planning, there is an urgent need to develop a fast and accurate Quality Assurance (QA) tool to predict the dose distribution for freehand interstitial brachytherapy. Much effort has been devoted to reducing plan variations in HDR brachytherapy. For example, knowledge-based planning (KBP) models have been proven to improve plan quality and predict the dose of new patients on the basis of historical information from treatment plans, such as patient anatomical information, distances from the HRCTV to organs at risk (OAR), and other geometric features [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Despite some progress achieved [\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], KBP-based methods are highly dependent on the selection of handcrafted features and can only predict the DVH of OARs [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn recent years, deep learning-based methods have achieved significant success in dose prediction for external beam radiation therapy (EBRT), and some have been gradually applied to brachytherapy. Kandalan et al. implemented a 3D convolutional neural network (CNN) to predict the dose distribution for patients with prostate cancer treated with volumetric modulated arc therapy (VMAT) [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Duan et al. developed a 3D asymmetrical ResNeSt dose prediction model for patients with esophageal cancer with various prescription levels [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Zhang et al. employed a neural network to predict the D\u003csub\u003e2cc\u003c/sub\u003e/D\u003csub\u003e90\u003c/sub\u003e dosimetric index of each OAR for patients with cervical cancer undergoing needle insertion in HDR brachytherapy [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Ma et al. proposed a 3D deep CNN to predict the dose distribution of tandem and ovoid (T\u0026amp;O) treatment plans for cervical cancer brachytherapy [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAlthough CNN models have been used and have demonstrated superiority for dose prediction in EBRT and T\u0026amp;O brachytherapy, there are still challenges in predicting the dose distribution for interstitial needle insertion. In particular, the locality of the receptive fields in the convolutional layers may lead to a poor ability to learn long-range spatial dependencies and global context [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]; thus, CNN models may yield suboptimal dose prediction for relatively small regions in the presence of large interpatient variation [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Consequently, it is challenging for CNN models to predict the dose distribution of freehand needle insertions in HDR brachytherapy, due to the complex geometric relationships between needles, such as crossing or touching. To address this issue, hybrid architectures that combine CNNs and self-attention mechanisms have been proposed to effectively enhance nonlocal modeling capabilities. Li et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] developed a deep neural network that combined the CNN model and the squeeze-and-excitation attention module to predict the 3D dose distribution in HDRBT, which demonstrated a higher level of agreement with clinical plans than the CNN model.\u003c/p\u003e \u003cp\u003eRecognizing the intricate nature of 3D dose prediction in HDRBT and the constraints of existing dose prediction methodologies, we propose an effective deep learning model for fast and accurate 3D dose prediction in HDR brachytherapy. Inspired by the tremendous success of the transformer mechanism [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], which is entirely based on the self-attention strategy and has a remarkable ability to capture the global context between sequential data, we developed the transformer mechanism as a feature encoder to overcome the limitations of CNN localized receptive fields. The purpose of this study is to investigate the feasibility of a 3D transformer-based deep learning model for dose prediction in HDR interstitial brachytherapy. The prediction performance is estimated via dose analysis, including the quantitative dose difference of Dose-Volume Histogram (DVH) metrics between clinical and predicted dose maps, 3D gamma analysis, the Dice similarity coefficient (DSC) of the 1\u0026ndash;30 Gy isodose volumes, and visualized dose distributions. To the best of our knowledge, this work is the first to employ transformer mechanisms to predict the 3D dose distribution of freehand HDR interstitial brachytherapy via needle insertion.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cp\u003eThe pipeline of the proposed method for predicting the voxel-wise dose distribution in HDR brachytherapy for cervical cancer patients is shown in Figure.1. The overall workflow includes (1) patient selection, (2) image processing, (3) network training, and (4) dosimetric evaluation and analysis. Each step is described in detail in sections 2.1 to 2.4.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Clinical data and treatment planning\u003c/h2\u003e \u003cp\u003eThis study retrospectively analyzed 96 CT-based treatment plans from 24 patients with different-stages of cervical cancer who were treated with freehand interstitial HDR brachytherapy via needle insertion. The patients\u0026rsquo; basic characteristics are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The average volume of the HRCTV was 94.6 cm\u003csup\u003e3\u003c/sup\u003e at the time of brachytherapy. Each patient was implanted with 4‒6 trocar stainless steel needles (Elekta AB, diameter of 1.5 mm, length of 200 mm) for a total of 4‒5 fractions of \u003csup\u003e192\u003c/sup\u003eIr HDR brachytherapy, and delivered with a dose of 6 Gy/fraction on the Flexitron HDR treatment afterloader (Elekta AB, Stockholm, Sweden). During brachytherapy, the radiation oncologist inserted trocar needles into the patient on the basis of the tumor topography, size, and OARs proximity. Subsequently, a CT scan was performed for each patient using a Siemens CT simulator (SOMATOM Sensation Open, Siemens Medical System, Germany) for contour delineation. All the CT images were scanned with a tube voltage of 120 kVp, a tube current of 150 mAs, a reconstructed size of 512\u0026times;512, and CT acquisition resolutions of 0.6\u0026ndash;1.2 mm in-plane and 5 mm longitudinal. The clinical target and OARs, including the high-risk clinical target volume (HRCTV), bladder, rectum, and bowel, were delineated by experienced radiation oncologists based on ESTRO recommendations [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. The CT images and contoured structures were then transferred to the Oncentra Brachy treatment planning system (Elekta-Brachy, Veenendaal, Netherlands) for brachytherapy treatment planning. The interstitial needles were manually digitized by experienced medical physicists to determine the trajectory of the radioactive source. The dose calculation algorithm was based on Task Group Report No.43, as recommended by the American Association of Physicists in Medicine (AAPM) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. The treatment planning for all patients was optimized via the Hybrid Inverse treatment Planning and Optimization algorithm (HIPO) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], and subsequently manually fine-tuned using graphical optimization tool. Finally, the clinical plan was approved by experienced radiation oncologists.\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\u003ePatient characteristics\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eCharacteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePatients\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e53.6\u0026thinsp;\u0026plusmn;\u0026thinsp;11.0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVolume of HR-CTV (cm\u003csup\u003e3\u003c/sup\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e94.6\u0026thinsp;\u0026plusmn;\u0026thinsp;54.3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePrescription dose (Gy) per fraction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6 Gy \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e 4\u0026thinsp;~\u0026thinsp;5 fx\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of needles (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.0\u0026thinsp;\u0026plusmn;\u0026thinsp;0.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eClinical stage (FIGO*)\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eⅠA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eⅡA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eⅡB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eⅢA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eⅢB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eⅢC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal patients\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal plans\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003e* FIGO (International Federation of Gynecology and Obstetrics)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Image processing\u003c/h2\u003e \u003cp\u003eThe overall treatment plan (N\u0026thinsp;=\u0026thinsp;96) was randomly divided into a training set (N\u0026thinsp;=\u0026thinsp;81) and a testing set (N\u0026thinsp;=\u0026thinsp;15). In this study, five structures, including the HRCTV, bladder, rectum, bowel, and needle trajectory, were converted into binary masks and combined as 3D input data. A zero-padding strategy was applied to the z-dimension of each input channel to maintain the same depth of 128. The corresponding dose distribution map with a grid size of 1 mm was considered the ground truth and aligned to the pixel spacing of the binary volumetric structure via trilinear interpolation.\u003c/p\u003e \u003cp\u003eIn addition, all the input images were cropped to 256\u0026times;256 pixels at the geometric center of the binary mask to eliminate unnecessary background pixels and address the GPU memory limitations. No data augmentation strategy was utilized because the high-dose distribution may follow the specific spatial relationship of the OARs and strict physical constraints. Finally, a volume size of 256\u0026times;256\u0026times;128\u0026times;5 was fed into the deep learning-based model for HDR brachytherapy dose prediction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Transformer-based CNN architecture\u003c/h2\u003e \u003cp\u003eInspired by the UNETR model [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], we reformulated the 3D HDR brachytherapy dose prediction task as a 1D sequence-to-sequence prediction problem. The network architecture of the proposed transformer-based CNN model is depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. We integrated a stack of transformer blocks into a 3D Unet-shaped deep learning network as a feature encoder to learn complex contextual information and long-range dependencies across 3D volumetric images. Subsequently, the feature maps from multiple resolutions of the transformer encoder were merged with the fine-grained feature maps in the upsampling CNN decoder via a skip-connection.\u003c/p\u003e \u003cp\u003eSpecifically, we first performed a patch partition layer to create a sequence of 3D tokens, and the flattened tokens were projected into a K-dimensional embedding space via a linear projection layer. To retain the spatial context of the extracted patch, 1D learnable positional embedding was added to the projected patch embedding. Thus, the input sequence Z\u003csub\u003e0\u003c/sub\u003e for the transformer encoder can be formulated as:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\begin{array}{c}{Z}_{0}=\\left[{x}_{p}^{1}E;{x}_{p}^{2}E;\\cdots\\:;{x}_{p}^{N}E\\right]+{E}_{pos}\\#\\left(1\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere x\u003csub\u003ep\u003c/sub\u003e\u0026isin;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\mathbb{R}}^{N\\times\\:({P}^{3}\\bullet\\:C)}\\)\u003c/span\u003e\u003c/span\u003e with each patch size of (p, p, p) and C input channels. N=(H\u0026times;W\u0026times;D)/P\u003csup\u003e3\u003c/sup\u003e represented the length of the sequence and (H, W, D) represented the resolution of the 3D input volume. E\u0026isin;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\mathbb{R}}^{({P}^{3}\\bullet\\:C)\\times\\:K}\\)\u003c/span\u003e\u003c/span\u003e represented the projected patch embedding and E\u003csub\u003epos\u003c/sub\u003e \u0026isin;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\mathbb{R}}^{N\\times\\:K}\\)\u003c/span\u003e\u003c/span\u003e represented the 1D position embedding. In this study, the 3D input volume (H\u0026times;W\u0026times;D\u0026times;C)=(256\u0026times;256\u0026times;128\u0026times;5), the patch resolution P\u0026thinsp;=\u0026thinsp;16, and the embedding size K\u0026thinsp;=\u0026thinsp;768.\u003c/p\u003e \u003cp\u003eAs the standard architecture in NLP, each transformer block consists of two main sublayers that follow layer normalization (LN): multihead self-attention (MSA) [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] and multilayer perceptron (MLP). A residual connection was added to each sublayer to enhance information flow. The encoder consists of several stacked transformer blocks, and the output Z\u003csub\u003ei\u003c/sub\u003e of the ith block in the transformer encoder can be described as follows:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\begin{array}{c}{Z}_{i}^{{\\prime\\:}}=MSA\\left(\\text{L}\\text{N}\\left({Z}_{i-1}\\right)\\right)+{Z}_{i-1}\\#\\left(2\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:\\begin{array}{c}{Z}_{i}=MLP\\left(\\text{L}\\text{N}\\left({Z}_{i}^{{\\prime\\:}}\\right)\\right)+{Z}_{i}^{{\\prime\\:}}\\#\\left(3\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{Z}_{i}^{{\\prime\\:}}\\)\u003c/span\u003e\u003c/span\u003e was the output of the MSA layer. i \u0026isin; {1, 2, \u0026hellip;\u0026hellip;, L} and L\u0026thinsp;=\u0026thinsp;12 was the total number of transformer blocks in the encoder.\u003c/p\u003e \u003cp\u003eFor the CNN decoder, the intermediate and bottleneck sequence representations Z\u003csub\u003ei\u003c/sub\u003e (i \u0026isin;{3,6,9,12}) from the transformer encoder were first reshaped to the input space (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{H}{p}\\times\\:\\frac{W}{p}\\times\\:\\frac{D}{p}\\times\\:K\\)\u003c/span\u003e\u003c/span\u003e), and the resolutions were increased by deconvolutional layers with a factor of 2. Furthermore, the transformed feature maps at multiresolutions were combined with a deconvolution layer followed by a 3\u0026times;3\u0026times;3 convolutional layer, batch normalization (BN), and rectified linear unit (ReLU) activation functions to adjust the latent dimension. The processed feature maps were then connected to the CNN-based decoder via skip connections. The concatenated feature maps were fed into a decoder block that contained two 3\u0026times;3\u0026times;3 convolutional layers, followed by a BN and ReLU, and a deconvolutional layer with a stride of 2 for upsampling to the input resolution. Notably, the decoder employed 6 additional convolutional layers to smoothly decrease the number of filters from 64 to 1 for better preservation of the extracted features [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], followed by a final 1\u0026times;1\u0026times;1 convolution layer to generate the predicted dose map.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Loss function and implementation details\u003c/h2\u003e \u003cp\u003eConsidering the high dose with steep gradients of dose distribution in HDR brachytherapy [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], we employed a voxelwise mean square error (MSE) as the loss function. For precise dose prediction, we also calculated the MSE loss in each ROI, including the body, HRCTV, bladder, rectum, bowel, and needle trajectories. The entire loss function of our proposed transformer-based CNN is expressed as:\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:\\begin{array}{c}{L}_{ROI}=\\frac{1}{n}\\sum\\:_{i=1}^{n}({\\text{Y}}_{i}-{\\widehat{Y}}_{i}{)}^{2}\\#\\left(4\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:\\begin{array}{c}Loss={\\lambda\\:}_{1}{L}_{body}+{\\lambda\\:}_{2}{L}_{HRCTV}+{\\lambda\\:}_{3}{L}_{bladder}+{\\lambda\\:}_{4}{L}_{rectum}+{\\lambda\\:}_{5}{L}_{bowel}+{\\lambda\\:}_{6}{L}_{needles}\\#\\left(5\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{Y}}_{i}\\)\u003c/span\u003e\u003c/span\u003e denoted the ground truth dose value, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\widehat{Y}}_{i}\\)\u003c/span\u003e\u003c/span\u003e denoted the predicted dose value, and n denoted the number of ROI points. λ\u003csub\u003ei\u003c/sub\u003e denoted the hyperparameters that balanced the five ROIs. This weighting factor allowed the network to place greater emphasis on learning the challenging regions of the dose prediction task. The coefficients λ1\u0026thinsp;=\u0026thinsp;λ2\u0026thinsp;=\u0026thinsp;3, and λ3\u0026thinsp;=\u0026thinsp;λ4\u0026thinsp;=\u0026thinsp;λ5\u0026thinsp;=\u0026thinsp;λ6\u0026thinsp;=\u0026thinsp;1 were considered.\u003c/p\u003e \u003cp\u003eThe proposed network was implemented using the PyTorch framework on an Intel Xeon Bronze 3206R CPU @ 1.90 GHz, NVIDIA Quadro RTX 6000 with 24 GB of memory. In addition, we employed the adaptive moment estimation (Adam) optimizer with a learning rate of 1\u0026times;10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e to minimize the MSE loss function. We trained the neural network for 500 epochs with an early stopping strategy, and the batch size was set to 1.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Dosimetric evaluation and statistical analysis\u003c/h2\u003e \u003cp\u003eFor dosimetric analysis, the mean error (ME) of the dosimetric parameters between the clinical and predicted dose maps was used to assess the prediction accuracy (ME\u0026thinsp;=\u0026thinsp;clinical-prediction). We compared the DVH parameters D\u003csub\u003e90\u003c/sub\u003e and D\u003csub\u003e95\u003c/sub\u003e for the HRCTV, D\u003csub\u003e1cc\u003c/sub\u003e, and D\u003csub\u003e2cc\u003c/sub\u003e for the OARs. The dose volume reporting parameters of the ESTRO recommendations [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] (V\u003csub\u003e100\u003c/sub\u003e, V\u003csub\u003e150\u003c/sub\u003e, and V\u003csub\u003e200\u003c/sub\u003e of the HRCTV) were used to calculate the dose distribution metrics to characterize the dose coverage, including the coverage index (CI\u0026thinsp;=\u0026thinsp;V\u003csub\u003e100\u003c/sub\u003e/V\u003csub\u003epres\u003c/sub\u003e), homogeneity index (HI\u0026thinsp;=\u0026thinsp;1-V\u003csub\u003e150\u003c/sub\u003e/V\u003csub\u003e100\u003c/sub\u003e), overdose volume index (ODI\u0026thinsp;=\u0026thinsp;V\u003csub\u003e200\u003c/sub\u003e/V\u003csub\u003e100\u003c/sub\u003e) and dose nonuniformity ratio (DNR\u0026thinsp;=\u0026thinsp;V\u003csub\u003e150\u003c/sub\u003e/V\u003csub\u003e100\u003c/sub\u003e) [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. D\u003csub\u003ex\u003c/sub\u003e (Gy) represents the dose received by x% of the HRCTV. D\u003csub\u003excc\u003c/sub\u003e represents the dose to the most exposed x cc volume of the OARs. V\u003csub\u003ex\u003c/sub\u003e represents the percentage volume of the corresponding ROI that receives at least x% of the prescribed dose. V\u003csub\u003epres\u003c/sub\u003e represents the volume of the prescribed dose region.\u003c/p\u003e \u003cp\u003eIn addition, a global 3D gamma analysis under the 3%/3 mm criterion was used to further evaluate the prediction accuracy. We calculated the Dice similarity coefficient (DSC) from 1 to 30 Gy to evaluate the isodose volume similarity between the transformer-based CNN model prediction and the clinical ground truth. The spatial dose distributions and DVH comparisons were visualized for qualitative evaluation. Furthermore, statistical analysis using paired t-tests was performed to assess the differences in dosimetric parameters between the network prediction and the clinical plan. A p-value less than 0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003eThe transformer-based CNN network took approximately 14 h to train a total of 400 epochs, and 8.5 s to predict a 3D dose distribution for cervical cancer brachytherapy.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e tabulated the mean and standard deviation of the dose differences for all the test sets. The 3D transformer-based CNN network achieved good dose agreement with the clinical treatment plans, with average prediction errors of D\u003csub\u003e90\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;0.08\u0026thinsp;\u0026plusmn;\u0026thinsp;0.32 Gy (p\u0026thinsp;=\u0026thinsp;0.41) and D\u003csub\u003e95\u003c/sub\u003e=-0.03\u0026thinsp;\u0026plusmn;\u0026thinsp;0.34 Gy (p\u0026thinsp;=\u0026thinsp;0.78) for the HRCTV. The mean errors of D\u003csub\u003e1cc\u003c/sub\u003e and D\u003csub\u003e2cc\u003c/sub\u003e for all OARs (bladder, rectum, and bowel) were less than 0.6 Gy. In particular, the mean errors of D\u003csub\u003e1cc\u003c/sub\u003e and D\u003csub\u003e2cc\u003c/sub\u003e for the bowel were less than 0.2 Gy. Moreover, the 3D gamma evaluation under the 3%/3 mm criterion was 92.73\u0026thinsp;\u0026plusmn;\u0026thinsp;3.82%, with the following dose distribution metrics: CI=-0.02\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05, HI=-0.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08, ODI\u0026thinsp;=\u0026thinsp;0.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11 and DNR\u0026thinsp;=\u0026thinsp;0.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08. Overall, there was no statistically significant difference in these dosimetric metrics between the deep learning predicted dose map and the clinical ground truth (p\u0026thinsp;\u0026gt;\u0026thinsp;0.05), except for D\u003csub\u003e1cc\u003c/sub\u003e and D\u003csub\u003e2cc\u003c/sub\u003e of the rectum. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e showed that the dose coverage of the HRCTV and the dose falling outside the target volume in the predicted dose map are highly consistent with clinical dose map.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a) and (b) showed an axial view of the clinical and predicted dose maps for the two test cases, with the corresponding DVHs and isodose lines depicted. The DVHs of the predicted dose maps matched well with the clinical ground truth. Additionally, the violin plot of different DVH metrics was presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(c) to find out the distribution of dose differences for each metric.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e showed the DSC calculation for evaluating the isodose volume similarity (1\u0026ndash;30 Gy) among all the testing cases. The average DSC of all testing cases for the 1\u0026ndash;30 Gy isodose level was 0.87 with a range of 0.77\u0026ndash;0.94, indicating high dose conformity between the predicted dose distribution and that in clinical practice.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDose differences (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD) of different dosimetric parameters of all the test cases.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\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=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDosimetric index\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3D transformer-based CNN models\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHRCTV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD\u003csub\u003e90\u003c/sub\u003e (Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.08\u0026thinsp;\u0026plusmn;\u0026thinsp;0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD\u003csub\u003e95\u003c/sub\u003e (Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e-0.03\u0026thinsp;\u0026plusmn;\u0026thinsp;0.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBladder\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD\u003csub\u003e1cc\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.6\u0026thinsp;\u0026plusmn;\u0026thinsp;1.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.27\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD\u003csub\u003e2cc\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.54\u0026thinsp;\u0026plusmn;\u0026thinsp;0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.16\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRectum\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD\u003csub\u003e1cc\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.58\u0026thinsp;\u0026plusmn;\u0026thinsp;0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.01\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD\u003csub\u003e2cc\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.57\u0026thinsp;\u0026plusmn;\u0026thinsp;0.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.01\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBowel\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD\u003csub\u003e1cc\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.16\u0026thinsp;\u0026plusmn;\u0026thinsp;1.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD\u003csub\u003e2cc\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.19\u0026thinsp;\u0026plusmn;\u0026thinsp;0.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eDose distribution metrics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e-0.02\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.17\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e-0.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.09\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eODI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.24\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDNR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.09\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGamma pass rate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3%/3mm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e92.73\u0026thinsp;\u0026plusmn;\u0026thinsp;3.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e/\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn this study, we proposed a 3D transformer-based deep learning framework to predict accurate dose distributions for cervical cancer HDR brachytherapy. Taking advantage of the hybrid CNN-transformer architecture, our proposed network not only encodes a strong global context and learns long-range dependencies via the transformer mechanism, but also effectively leverages high-resolution spatial details via CNNs. Furthermore, we developed an MSE loss function that supervised the dose accuracy within body and region of interest (ROI) to improve network performance. Our proposed network generated a real-time dose map based on patient anatomy and needle geometry, which has the clinical potential to help oncologists adjust interstitial implants to avoid an insufficient dose to the target volume and provide better protection of the OARs. This approach can serve as an automatic QA tool in clinical practice to reduce human error during the operation process and further improve treatment efficacy.\u003c/p\u003e \u003cp\u003eDose prediction in interstitial brachytherapy is a more complex process than in EBRT, as its dose distribution is highly constrained by patient anatomy and insertion geometry. Highly inhomogeneous dose distributions and steep dose gradients within the HRCTV are great challenges that hamper the development of automatic dose prediction for HDRBT. Although few studies have investigated deep learning-based methods to facilitate 3D dose prediction for HDR brachytherapy, they have focused mostly on tandem and ovoid (T\u0026amp;O) or tandem-and-ring (T\u0026amp;R) applicators, with little attention given to free-hand interstitial needles. Previously, Cortes et al. developed a 3D CNN to predict dose maps for T\u0026amp;O brachytherapy and reported a mean dose difference of -0.09\u0026thinsp;\u0026plusmn;\u0026thinsp;0.67 Gy for HRCTV D\u003csub\u003e90\u003c/sub\u003e, -0.17\u0026thinsp;\u0026plusmn;\u0026thinsp;0.67 Gy for bladder D\u003csub\u003e2cc\u003c/sub\u003e, and \u0026minus;\u0026thinsp;0.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.46 Gy for rectum D\u003csub\u003e2cc\u003c/sub\u003e [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Ma et al. employed a 3D CNN to predict 3D dose distributions for T\u0026amp;O brachytherapy and proposed a dose difference of 1% for CTV V\u003csub\u003e100\u003c/sub\u003e, 0.2 Gy for bladder D\u003csub\u003e2cc\u003c/sub\u003e, 0.25 Gy for rectum D\u003csub\u003e2cc\u003c/sub\u003e and 0.11 Gy for bowel D\u003csub\u003e2cc\u003c/sub\u003e [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Li et al. demonstrated an attention-based deep learning method for 3D brachytherapy dose prediction and predicted average dose differences of 0.37\u0026thinsp;\u0026plusmn;\u0026thinsp;0.25 Gy for HRCTV D\u003csub\u003e90\u003c/sub\u003e, 0.23\u0026thinsp;\u0026plusmn;\u0026thinsp;0.14 Gy for bladder D\u003csub\u003e2cc\u003c/sub\u003e, and 0.28\u0026thinsp;\u0026plusmn;\u0026thinsp;0.20 Gy for rectum D\u003csub\u003e2cc\u003c/sub\u003e [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eObviously, dose prediction for interstitial brachytherapy via needle insertion is a more challenging undertaking than for T\u0026amp;O brachytherapy. This is attributable to the increased complexity of the needle geometry, and the varying number of needles may further complicate the dose distribution within the HRCTV. Comparatively, our proposed 3D transformer-based methods presented mean dose differences of 0.08\u0026thinsp;\u0026plusmn;\u0026thinsp;0.32 Gy for HRCTV D\u003csub\u003e90\u003c/sub\u003e, 0.54\u0026thinsp;\u0026plusmn;\u0026thinsp;0.89 Gy for bladder D\u003csub\u003e2cc\u003c/sub\u003e, 0.57\u0026thinsp;\u0026plusmn;\u0026thinsp;0.42 Gy for rectum D\u003csub\u003e2cc,\u003c/sub\u003e and 0.19\u0026thinsp;\u0026plusmn;\u0026thinsp;0.93 Gy for bowel D\u003csub\u003e2cc\u003c/sub\u003e (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Besides, to further evaluate the dose accuracy, gamma analysis was performed and achieved a 3%/3 mm gamma passing rate of 92.73\u0026thinsp;\u0026plusmn;\u0026thinsp;3.82% between the network prediction and the clinical ground truth. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e (a), visualization of the 1\u0026ndash;30 Gy isodose level revealed that the predicted dose map inherits a similar dose distribution to the clinical dose map. High average DSC (\u0026gt;\u0026thinsp;0.75) between the predicted dose distribution and the clinical ground truth was observed at the 1\u0026ndash;30 Gy isodose level (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Our proposed transformer-based CNN method can predict 3D dose distributions consistent with clinical treatment planning.\u003c/p\u003e \u003cp\u003eThere are several limitations in our study. First, the entire transformer-based network is designed for 3D volume to better preserve the specific spatial relationships between slices. However, this resulted in a notable increase in the computational intensity of the GPU. To address this, network hyperparameters such as batch size, the number of transformer blocks, and the channel of convolutional layers were adjusted to reduce the computational complexity, which may lead to suboptimal prediction in certain cases. Second, the proposed network still requires patient contours as input and may present potential uncertainty in manual delineation. Third, in accordance with the findings of Kallis et al. [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], a larger dose difference (isodose DSC of 0.77) was observed at a small dose level (1 Gy), which may be due to the steep dose gradients caused by the single dwell positions. Consequently, more clinical plan data need to be employed to achieve better training of the deep learning model. Moreover, considering that only one institution was included in this project and the varying treatment protocols of different institutions, multi-institutional involvement is essential for the development of a robust and accurate dose prediction model.\u003c/p\u003e \u003cp\u003eThis study is tailored specifically to fully interstitial situations. In future work, the proposed approach can be expanded to facilitate hybrid applicator scenarios to develop clinical benefits. The dwell time and dwell position of the radioactive source based on the predicted dose distribution can be further calculated to improve the automation of the HDR brachytherapy treatment process.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eIn this work, we proposed and evaluated a 3D transformer-based CNN method that can serve as a QA tool for needle insertion in HDR interstitial brachytherapy. Dosimetric evaluation showed that the predicted dose distributions were in good agreement with the clinical dose plans. This method has the clinical potential to facilitate brachytherapy treatment workflow and standardize brachytherapy treatment planning.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eHDRBT High-dose-rate brachytherapy\u003c/p\u003e\n\u003cp\u003eEBRT External beam radiation therapy\u003c/p\u003e\n\u003cp\u003eHRCTV High risk clinical target volume\u003c/p\u003e\n\u003cp\u003eOARs Organs at risk\u003c/p\u003e\n\u003cp\u003eT\u0026amp;O Tandem and ovoid\u003c/p\u003e\n\u003cp\u003eT\u0026amp;R Tandem-and-ring\u003c/p\u003e\n\u003cp\u003eFIGO International Federation of Gynecology and Obstetrics\u003c/p\u003e\n\u003cp\u003eKBP knowledge-based planning\u003c/p\u003e\n\u003cp\u003eCNN Convolutional neural network\u003c/p\u003e\n\u003cp\u003eHIPO Hybrid Inverse treatment Planning and Optimization algorithm\u003c/p\u003e\n\u003cp\u003eQA Quality assurance\u003c/p\u003e\n\u003cp\u003eDSC Dice similarity coefficient\u003c/p\u003e\n\u003cp\u003eDVH Dose-Volume Histogram\u003c/p\u003e\n\u003cp\u003eCI Coverage index\u003c/p\u003e\n\u003cp\u003eHI Homogeneity index\u003c/p\u003e\n\u003cp\u003eODI Overdose volume index\u003c/p\u003e\n\u003cp\u003eDNR Dose nonuniformity ratio\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThis study was reviewed and approved by the Institutional Review Board of the Second Affiliated Hospital of Guangzhou University of Chinese Medicine (No. ZF2020-084.2). It was conducted in accordance with the Declaration of Helsinki 1964 and its later amendments or comparable ethical standards. Consent was waived due to the retrospective nature of this study.\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThe authors declare that they have no competing interests\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis study is supported by grants from the National Natural Science Foundation of China (Grant No. 82172020), the State Key Laboratory of Traditional Chinese Medicine Syndrome (Grant No. QZ2023ZZ05), and the Ministry of Education Industry-Academic Cooperation Project (Grant No. 230806071142911).\u003c/p\u003e\n\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\n\u003cp\u003eWeiwei Guo and Wanwei Jian performed conceptualization, methodology, and writing of the original draft. Lin Zhu and Bailin Zhang contributed data curation and validation. Qiang He performed validation, reviewing and editing of the manuscript. Geng Yang supervised the study. Xuetao Wang contributed to the conceptualization and project administration. All the authors have read and approved the final manuscript. Weiwei Guo and Wanwei Jian contributed equally as the first authors.\u003c/p\u003e\n\u003ch2\u003eAcknowledgements\u003c/h2\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eThe datasets used during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSturdza A, P\u0026ouml;tter R, Fokdal LU, Haie-Meder C, Tan LT, Mazeron R, et al. Image guided brachytherapy in locally advanced cervical cancer: Improved pelvic control and survival in RetroEMBRACE, a multicenter cohort study. Radiotherapy Oncology: J Eur Soc Therapeutic Radiol Oncol. 2016;120:428\u0026ndash;33. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.radonc.2016.03.011\u003c/span\u003e\u003cspan address=\"10.1016/j.radonc.2016.03.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChargari C, Deutsch E, Blanchard P, Gouy S, Martelli H, Gu\u0026eacute;rin F, et al. Brachytherapy: An overview for clinicians. Cancer J Clin. 2019;69:386\u0026ndash;401. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3322/caac.21578\u003c/span\u003e\u003cspan address=\"10.3322/caac.21578\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNomden CN, de Leeuw AAC, Moerland MA, Roesink JM, Tersteeg RJHA, J\u0026uuml;rgenliemk-Schulz IM. Clinical use of the Utrecht applicator for combined intracavitary/interstitial brachytherapy treatment in locally advanced cervical cancer. Int J Radiat Oncol Biol Phys. 2012;82:1424\u0026ndash;30. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ijrobp.2011.04.044\u003c/span\u003e\u003cspan address=\"10.1016/j.ijrobp.2011.04.044\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNomden CN, de Leeuw AAC, Moerland MA, Roesink JM, Tersteeg RJHA, J\u0026uuml;rgenliemk-Schulz IM. Clinical use of the Utrecht applicator for combined intracavitary/interstitial brachytherapy treatment in locally advanced cervical cancer. Int J Radiat Oncol Biol Phys. 2012;82:1424\u0026ndash;30. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ijrobp.2011.04.044\u003c/span\u003e\u003cspan address=\"10.1016/j.ijrobp.2011.04.044\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGe Y, Wu QJ. Knowledge-based planning for intensity-modulated radiation therapy: A review of data-driven approaches. Med Phys. 2019;46:2760\u0026ndash;75. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.13526\u003c/span\u003e\u003cspan address=\"10.1002/mp.13526\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCortes KG, Kallis K, Simon A, Mayadev J, Meyers SM, Moore KL. Knowledge-based three-dimensional dose prediction for tandem-and-ovoid brachytherapy. Brachytherapy. 2022;21:532\u0026ndash;42. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.brachy.2022.03.002\u003c/span\u003e\u003cspan address=\"10.1016/j.brachy.2022.03.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKallis K, Mayadev J, Kisling K, Brown D, Scanderbeg D, Ray X, et al. Knowledge-based dose prediction models to inform gynecologic brachytherapy needle supplementation for locally advanced cervical cancer. Brachytherapy. 2021;20:1187\u0026ndash;99. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.brachy.2021.07.001\u003c/span\u003e\u003cspan address=\"10.1016/j.brachy.2021.07.001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYusufaly TI, Kallis K, Simon A, Mayadev J, Yashar CM, Einck JP, et al. A knowledge-based organ dose prediction tool for brachytherapy treatment planning of patients with cervical cancer. Brachytherapy. 2020;19:624\u0026ndash;34. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.brachy.2020.04.008\u003c/span\u003e\u003cspan address=\"10.1016/j.brachy.2020.04.008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWen L, Xiao J, Zeng J, Zu C, Wu X, Zhou J, et al. Multi-level progressive transfer learning for cervical cancer dose prediction. Pattern Recogn. 2023;141:109606. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.patcog.2023.109606\u003c/span\u003e\u003cspan address=\"10.1016/j.patcog.2023.109606\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKandalan RN, Nguyen D, Rezaeian NH, Barrag\u0026aacute;n-Montero AM, Breedveld S, Namuduri K, et al. Dose prediction with deep learning for prostate cancer radiation therapy: Model adaptation to different treatment planning practices. Radiotherapy Oncology: J Eur Soc Therapeutic Radiol Oncol. 2020;153:228\u0026ndash;35. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.radonc.2020.10.027\u003c/span\u003e\u003cspan address=\"10.1016/j.radonc.2020.10.027\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuan Y, Wang J, Wu P, Shao Y, Chen H, Wang H, et al. AS-NeSt: A Novel 3D Deep Learning Model for Radiation Therapy Dose Distribution Prediction in Esophageal Cancer Treatment With Multiple Prescriptions. Int J Radiation Oncology*Biology*Physics. 2024;119:978\u0026ndash;89. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.ijrobp.2023.12.001\u003c/span\u003e\u003cspan address=\"10.1016/j.ijrobp.2023.12.001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang H-W, Zhong X-M, Zhang Z-H, Pang H-W. Dose prediction of organs at risk in patients with cervical cancer receiving brachytherapy using needle insertion based on a neural network method. BMC Cancer. 2023;23:385. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12885-023-10875-6\u003c/span\u003e\u003cspan address=\"10.1186/s12885-023-10875-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa M, Kidd E, Fahimian BP, Han B, Niedermayr TR, Hristov D, et al. Dose Prediction for Cervical Cancer Brachytherapy Using 3-D Deep Convolutional Neural Network. IEEE Trans Radiation Plasma Med Sci. 2022;6:214\u0026ndash;21. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/TRPMS.2021.3098507\u003c/span\u003e\u003cspan address=\"10.1109/TRPMS.2021.3098507\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen J, Lu Y, Yu Q, Luo X, Adeli E, Wang Y et al. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation. ArXiv. 2021; abs/2102.04306.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHatamizadeh A, Yang D, Roth HR, Xu D. UNETR: Transformers for 3D Medical Image Segmentation. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2021:1748\u0026ndash;58.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Z, Yang Z, Lu J, Zhu Q, Wang Y, Zhao M, et al. Deep learning-based dose map prediction for high-dose-rate brachytherapy. Phys Med Biol. 2023;68. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1361-6560/acecd2\u003c/span\u003e\u003cspan address=\"10.1088/1361-6560/acecd2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVaswani A, Shazeer NM, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is All you Need. Neural Information Processing Systems; 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu W, Jiang J, Tian H, Gu J, Lu Y, Yang W, et al. Asymmetric convolution Swin transformer for medical image super-resolution. Alexandria Eng J. 2023;85:177\u0026ndash;84. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.aej.2023.11.044\u003c/span\u003e\u003cspan address=\"10.1016/j.aej.2023.11.044\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eP\u0026ouml;tter R, Haie-Meder C, Van Limbergen E, Barillot I, De Brabandere M, Dimopoulos J, et al. Recommendations from gynaecological (GYN) GEC ESTRO working group (II): concepts and terms in 3D image-based treatment planning in cervix cancer brachytherapy-3D dose volume parameters and aspects of 3D image-based anatomy, radiation physics, radiobiolog. Radiotherapy and Oncology. J Eur Soc Therapeutic Radiol Oncol. 2006;78:67\u0026ndash;77. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.radonc.2005.11.014\u003c/span\u003e\u003cspan address=\"10.1016/j.radonc.2005.11.014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRivard MJ, Coursey BM, DeWerd LA, Hanson WF, Huq MS, Ibbott GS, et al. Update of AAPM Task Group 43 Report: A revised AAPM protocol for brachytherapy dose calculations. Med Phys. 2004;31:633\u0026ndash;74. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1118/1.1646040\u003c/span\u003e\u003cspan address=\"10.1118/1.1646040\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKarabis A, Giannouli S, Baltas D. 40 HIPO: A hybrid inverse treatment planning optimization algorithm in HDR brachytherapy. Radiother Oncol 2005;76.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ArXiv 2020;abs/2010.1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTanderup K, M\u0026eacute;nard C, Polgar C, Lindegaard JC, Kirisits C, P\u0026ouml;tter R. Advancements in brachytherapy. Adv Drug Deliv Rev. 2017;109:15\u0026ndash;25. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.addr.2016.09.002\u003c/span\u003e\u003cspan address=\"10.1016/j.addr.2016.09.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaur G, Garg P, Srivastava AK, Gaur G, Sheetal, Grover R, et al. Dosimetric and radiobiological evaluation of treatment plan for cervical cancer high-dose-rate intracavitary brachytherapy. J Contemp Brachytherapy. 2022;14:253\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5114/jcb.2022.117729\u003c/span\u003e\u003cspan address=\"10.5114/jcb.2022.117729\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKallis K, Moore LC, Cortes KG, Brown D, Mayadev J, Moore KL et al. Automated treatment planning framework for brachytherapy of cervical cancer using 3D dose predictions. Phys Med Biol 2023;68. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1361-6560/acc37c\u003c/span\u003e\u003cspan address=\"10.1088/1361-6560/acc37c\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-cancer","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bcan","sideBox":"Learn more about [BMC Cancer](http://bmccancer.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bcan/default.aspx","title":"BMC Cancer","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Brachytherapy, dose prediction, transformer, deep learning, cervical cancer","lastPublishedDoi":"10.21203/rs.3.rs-5069942/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5069942/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eThe clinical process of high-dose-rate brachytherapy (HDRBT) is time-consuming and reliant on user expertise and preference. Deep learning-based dose prediction can act as a quality assurance (QA) tool to identify suboptimal needle placement and enhance treatment planning efficiency in HDR interstitial brachytherapy. This study aims to introduce a 3D transformer-based deep learning method to automatically predict the dose distribution in HDR brachytherapy for cervical cancer.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe retrospectively analyzed 96 CT-based treatment plans from 24 cervical cancer patients who underwent interstitial HDRBT with needle insertion. The transformer mechanism was integrated into a convolutional neural network (CNN) to capture long-distance characteristics and global information. The prediction performance was evaluated by the mean error of Dose-Volume Histogram metrics between clinical and predicted dose maps, gamma analysis, and the Dice similarity coefficient (DSC) of the 1\u0026ndash;30 Gy isodose volumes.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe mean error of D\u003csub\u003e90\u003c/sub\u003e and D\u003csub\u003e95\u003c/sub\u003e for HRCTV was less than 0.1 Gy, and the mean D\u003csub\u003e2cc\u003c/sub\u003e for organs at risk (bladder, rectum, and bowel) was under 0.6 Gy. The mean DSC of the 1\u0026ndash;30 Gy isodose volume was 0.87. The 3D transformer-based CNN model can predict dose maps that are highly consistent with the clinical treatment plans.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eA novel 3D transformer-based deep learning model was successfully developed for dose prediction in HDR interstitial brachytherapy. This method can automatically generate accurate 3D dose distributions, exhibiting great clinical potential for improving treatment efficiency and standardizing brachytherapy treatment planning.\u003c/p\u003e\u003ch2\u003eTrial registration:\u003c/h2\u003e \u003cp\u003eNo. ZF2020-084.2; Oct 16, 2020\u003c/p\u003e","manuscriptTitle":"3D transformer-based dose prediction in HDR brachytherapy for cervical cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-09 13:01:09","doi":"10.21203/rs.3.rs-5069942/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorAssigned","content":"","date":"2024-09-12T17:28:00+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-09-12T12:26:01+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Cancer","date":"2024-09-11T09:31:52+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-cancer","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bcan","sideBox":"Learn more about [BMC Cancer](http://bmccancer.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bcan/default.aspx","title":"BMC Cancer","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c23342da-986a-41bc-a804-649ea543d484","owner":[],"postedDate":"December 9th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-01-26T16:02:18+00:00","versionOfRecord":{"articleIdentity":"rs-5069942","link":"https://doi.org/10.1186/s12885-026-15592-4","journal":{"identity":"bmc-cancer","isVorOnly":false,"title":"BMC Cancer"},"publishedOn":"2026-01-20 15:57:22","publishedOnDateReadable":"January 20th, 2026"},"versionCreatedAt":"2024-12-09 13:01:09","video":"","vorDoi":"10.1186/s12885-026-15592-4","vorDoiUrl":"https://doi.org/10.1186/s12885-026-15592-4","workflowStages":[]},"version":"v1","identity":"rs-5069942","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5069942","identity":"rs-5069942","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}