Multi-Reader Evaluation of Deep Learning-Based Auto-Segmentation of Eloquent Brain Arteriovenous Malformation on MRA and White Matter Tractography in Stereotactic Radiosurgery

Multi-Reader Evaluation of Deep Learning-Based Auto-Segmentation of Eloquent Brain Arteriovenous Malformation on MRA and White Matter Tractography in Stereotactic Radiosurgery | 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 Multi-Reader Evaluation of Deep Learning-Based Auto-Segmentation of Eloquent Brain Arteriovenous Malformation on MRA and White Matter Tractography in Stereotactic Radiosurgery Mingzhu Li, Huaguang Zhu, Yun Guan, Xiaoxia Liu, Xiaojia Gong, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6632250/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 02 Mar, 2026 Read the published version in Radiation Oncology → Version 1 posted 9 You are reading this latest preprint version Abstract Objective: To minimize the radiation injury for white matter (WM) pathways during brain arteriovenous malformation (bAVM) stereotactic radiosurgery (SRS), the WM tractography is integrated into treatment planning to identify WM pathways and restrict receiving dose. Manual segmentation of eloquent bAVM adjacent to WM pathways is time-consuming and prone to substantial inter-practitioner variability due to intricate entanglement within eloquent brain areas. The objective of this study is to develop and evaluate a deep learning (DL) system for the segmentation of eloquent bAVM in a clinical setting. Methods: A total of 191 eloquent bAVM patients who underwent WM tractography and 3D time-of-flight magnetic resonance angiography (TOF-MRA) images were enrolled. 153 patients were used to construct a two-stage DL bAVM segmentation ensemble (TBASE) consisting of 2D detection and 3D segmentation models to segment the bAVM, the other 38 to test performance. Comparative experiments with Res-Net and U-Net were conducted to validate the effectiveness of the proposed network. A randomized multi-reader evaluation was then conducted to assess the impact of TBASE assistance for bAVM segmentation using ten algorithm-unseen cases. Six medical professionals contoured the same series of cases in both assisted and unassisted modes, with a 6-week memory washout period between each session. The aided and unaided Dice Similarity Coefficients (DSC), Hausdorff Distance (HD), along with contouring times were compared. Results: The mean values and standard deviations for DSC and HD of TBASE are 0.87 ± 0.03 and 3.51 ± 0.26, respectively, while Res-Net and U-Net results are 0.75 ± 0.12 and 4.14 ± 0.99, 0.77 ± 0.09 and 3.94 ± 0.82, respectively. The average volume difference across all patients in test dataset is 0.25 ± 1.39 cc, with no statistically significant variation observed. With TBASE assistance, the mean DSC of readers improved from 0.76 ± 0.07 to 0.86 ± 0.05 ( P < 0.001), with corresponding values of mean HD reducing from 4.31 ± 0.68 to 3.35 ± 0.17 ( P < 0.001) and a mean time saving of 52.15% ± 13.85% per patient. Less-experienced readers achieved greater improvements in contouring accuracy compared to specialists (DSC increase: 0.15 ± 0.11 vs 0.06 ± 0.09; P < 0.001), while demonstrating similar reductions in contouring time as specialists (50.68% ± 14.62% vs. 55.1% ± 11.98%; P = 0.217) Conclusions: The reliable eloquent bAVM automated-segmentation method has been validated in clinical workflow. The TBASE assistance improved the accuracy and efficiency of the eloquent bAVM manual delineation and thus avoid neurological sequelae after SRS in considering WM pathways protection. Brain arteriovenous malformation White matter tractography Stereotactic radiosurgery Deep learning Automatic segmentation Randomized multi-reader Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction Brain arteriovenous malformation is a vascular anomaly typically characterized by abnormal connections between arteries and veins, potentially leading to disrupted blood flow and increased risk of bleeding 1 , 2 . Stereotactic radiosurgery, including technologies like CyberKnife, has been widely recognized as an effective therapeutic approach for managing bAVM 3 , 4 . Due to the therapeutic benefits of delivering maximal radiation dose, a crucial aspect of radiosurgery is determining the optimal radiation margin while safeguarding language, visual, and motor functions 5 , 6 . Hence, accurately localizing eloquent cortical regions and WM pathways within the bAVM margin is imperative for avoiding neurological sequelae 7 . Recently, WM tractography has been incorporated into the treatment planning process for SRS 8 – 10 . This integration aided in identifying the anatomical relationship between WM fibers and nidus, thereby facilitating the optimization of treatment plan design 11 , 12 . Accurate identification and delineation of the bAVM nidus are essential stages in SRS treatment planning. Presently, targets are manually identified by neuroradiologists and contoured by radiation oncologists, utilizing multi-modal and WM Tractography images while referring to digital subtraction angiography images. However, this entire process is often time-consuming, labor-intensive, and susceptible to inter-observer variability, especially when dealing with locations adjacent to WM pathways. With DL algorithms represent the forefront of medical image analysis 13 – 16 , auto-segmentation is devised to address the limitations of bAVM’s manual contours by delivering faster and observer-independent results. A U-Net model was proposed by Jiao et al. to detect and quantify the diffuseness of bAVM nidus in MRA image 17 . Wang et al. developed a deeply supervised V-Net to delineate the bAVM nidus from CT images and evaluated target dose coverage changes on original plan 18 . Both studies have validated the reliability of using DL for the direct segmentation of bAVM nidus on medical images. However, there are still a lot of limitations. These methodologies predominantly rely on single-modal images, which may not adequately capture the complex morphology and characteristics of bAVM. Furthermore, lesions located near WM pathways often received radiation doses exceeding safe thresholds due to difficulties in their precise identification on CT or conventional MRI scans. Lastly, despite the promising standalone performance of DL in preclinical evaluations of bAVM SRS, there has been a significant lack of clinical trials to evaluate how these results translate into practical clinical settings. To fully assess the clinical potential of DL in managing bAVM, it is crucial to incorporate the system into clinical practice and evaluate its effect on practitioners. Given the limitations mentioned above, we have developed an innovative two-stage DL ensemble (TBASE), which integrated 2D U-Net detection-aided and 3D CMUnet-Cbam for automatic bAVM segmentation. This approach leverage multi-modal imaging alongside WM tractography, marking the first utilization of WM tractography in the auto-segmentation of bAVM. It aims to enhance the precision of target data extraction, thus ensuring the preservation of WM pathways. The results of bAVM auto-segmentations are evaluated for standard geometric indices including DSC and HD. Subsequently, we conducted a cross-modal, multi-reader study involving physicians across various SRS experience levels and specialties, utilizing TBASE for target contouring of eloquent bAVM. We assessed inter-reader variability, accuracy, and the efficiency of assisted contouring for eloquent bAVM SRS. 2. Methods and Materials 2.1 Clinical dataset A retrospective analysis is conducted on 191 patients with bAVM located either adjacent to or within WM pathways, all of whom underwent CyberKnife radiosurgery at Huashan Hospital, Fudan University, from 2016 and 2023. Our retrospective study was approved by the Ethical Review Committee of the Huashan Hospital, Fudan University and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived owing to the retrospective nature of the study. Demographic and clinical characteristics, including age, sex, lesion size, affected hemisphere, specific lesion location, and Spetzler-Martin grading 19 are documented in Table 1 . Statistical analysis revealed no significant differences in these baseline characteristics between the training and testing sets utilized in the segmentation model. All manual contours are rigorously identified by neuroradiologists and contoured by radiation oncologists to ensure quality control. Image acquisition information is shown in Supplementary Appendix A. Table 1 Demographic and clinical characteristics of patients for bAVM segmentation. Training Set (n = 153) Test Set (n = 38) Age, median [IQR] 26 [17–62] 25 [19–48] Gender, number (%) Female 66 (43.1%) 18 (47.4%) Male 87 (56.9%) 20 (52.6%) Size (cc), median [IQR] 15.4 [1.3–55.9] 15.1 [1.5–48.6] Lesion-affected hemisphere. number (%) Left 74 (48.4%) 18 (47.4%) Right 79 (51.6%) 20 (52.6%) Lesion location, number (%) Frontal lobe 51 (33.3%) 13 (34.2%) Parietal lobe 29 (19%) 12 (31.6%) Basal ganglia 73 (47.8%) 13 (34.2%) S-M grading, number (%) II 61 (39.9%) 14 (36.8%) III 84 (54.9%) 23 (60.5%) IV 8 (5.2%) 1 (2.6%) IQR interquartile range, S-M grading Spetzler-Martin grading. 2.2 Data preprocessing To construct TBASE, the 153 patients are divided into training and validation group with a ratio of approximately 80%: 20% (122: 31), and the other 38 patients are reserved to test performance according to different locations (frontal lobe, parietal lobe and basal ganglia). Image registration of each patient’s WM tractography to their corresponding MRA image is accomplished using a proprietary MIM workflow 20 . To enhance the robustness of the model, data augmentation such as rotation, flipping, and zooming are employed, facilitating the model's ability to learn transformation-invariant features. 2.3 Segmentation 2.3.1 Overview A two-stage deep learning ensemble TBASE is trained to segment bAVM nidus, which consisting of a 2D U-Net detection model to localize the bAVM ROI and 3D CMUnet-Cbam to segment bAVM within identified location. The whole processes for training and validation phase are illustrated in upper part of Fig. 1 . In the first stage of training phase, the multimodal images of each training patient are served as input images and concatenated as multi-channel inputs, with the manual contours serving as the target for learning. The binary mask of bAVM is coarsely generated in 2D U-Net and then a 128×128×128 pixels ROI bounding box centered at the mass center of mask is created to localize the bAVM nidus. The coordinates of these cropped ROIs are stored and subsequently utilized to map the predicted masks back to the original images. In the second stage, the multi-modal images are cropped to the defined ROI, and these images are then input into the 3D CMUnet-Cbam architecture, which provided accurate and efficient bAVM segmentation within the ROI. In the testing phase, for a new multi-modal image dataset, the localization of the bAVM ROI and subsequent segmentation within that ROI are achieved following the same process as in the training phase. The final segmented bAVM contour is then reconstructed by positioning the segmented bAVM ROI within the full 3D image volume. The 2D U-Net and 3D CMUnet-Cbam models are implemented using Python 3.8 using Pytorch 1.12, enhanced by CUDA 12.2, on six NVIDIA RTX 4080Ti GPUs. The models underwent training and testing across 300 epochs, employing the Adam optimizer, which incorporated a momentum term of 0.5 to enhance convergence. The initial learning rate is set at 0.0002 and systematically reduced by half whenever the error rates on the validation set plateaued, ensuring efficient and effective learning progression. 2.3.2 Network architecture The architecture of our proposed CMUnet-Cbam is detailed in bottom part of FIG.1, and it is divided into two stages: encoding and decoding. In the encoding stage, which includes five levels organized from top to bottom, the input image slices which cropped in the detection phase initially pass through a combination of conventional convolution and attention blocks, with max-pooling employed to reduce the size of the feature maps. And the following four levels utilize the CMUNeXt block paired with an attentional block. To maintain the multi-level features effectively, concatenation is employed to enhance the volume of feature maps from both the encoding and decoding stages. The final transformation involves funneling these maps through a convolution layer to decrease their dimensionality to two channels. This is immediately followed by a “tanh” activation layer that serves to polarize the feature maps, thus distinguishing the posterior probabilities of the bAVM nidus from normal tissue. The network specifications are described in detail in Supplementary Appendix B. To validate the effectiveness of the proposed network, we conducted comparative experiments with Res-Net and U-Net 21 , 22 . 2.4 Multi-reader study design The study employs a cross-modal, multi-reader design as illustrated in Fig. 2 . A separate dataset comprising of ten bAVM patients are randomly selected from test dataset for the reader study. All cases are independent of the development dataset used for model training. Two specialist and four non-specialist SRS physicians contoured the selected dataset cases, with or without TBASE assistance. To minimize potential bias from previous contouring, participants are randomly assigned to either the assisted-first or unassisted-first segmentation mode. The two contour sessions are separated by an interval of at least 6-weeks. During the second session, physicians are blinded to both their initial segmentations and those from other participants. For this separate dataset, the GT is independently delineated by two SRS experts. Three-dimensional reconstructions of their contours are simultaneously visualized and compared, allowing for direct assessment of inter-expert variability. Any discrepancies observed are resolved through a consensus process. 2.5 Evolution metrics and Statistical Analysis Quantitative comparisons between the segmentation results from our proposed model and manual contouring are conducted for each patient in the validation dataset and multi-reader study cases, employing metrics such as DSC and HD 23 , 24 . The metrics employed in this study are defined as follows: $$\:DSC(A,B)=\frac{2|A\cap\:B|}{\left|A\right|+\left|B\right|}$$ 1 $$\:HD(A,B)=\text{max}\left(\text{h}\left(A,B\right),\text{h}\left(B,A\right)\right)$$ 2 Where A and B represent auto-segment and manual contours, respectively. \(\:\text{h}\text{(}A,B\text{)}\text{=}{\text{max}}_{\text{a}\in\text{A}}{\text{min}}_{\text{b}\in\text{B}}\left|\right|\text{a}\text{-}\text{b||}\) , \(\:\text{h}\text{(}B,A\text{)}\text{=}{\text{max}}_{\text{b}\in\text{B}}{\text{min}}_{\text{a}\in\text{A}}\left|\right|\text{b}\text{-}\text{a}\text{||}\) , \(\:\left|\right|\text{.}\text{||}\) denotes the Euclidean distance. We conducted Wilcoxon paired-signed tests to compare the DSC, HD and time consumed between two contouring modes for bAVM detection. The statistical significance is determined with a two-tailed P < 0.05. Meanwhile, the Pearson correlation coefficient (PCC) is used to evaluate the influence of bAVM size on the DSC and HD metrics. 3. Results 3.1 TBASE Performance The overall geometric parameters for all 38 patients in test dataset and model evaluation are numerical summarized in FIG. 3 . As shown in Fig. 3 .A, the mean values and standard deviations for DSC and HD are 0.87 ± 0.03 and 3.51 ± 0.26, respectively, while Res-Net and U-Net results are 0.75 ± 0.12 and 4.14 ± 0.99, 0.77 ± 0.09 and 3.94 ± 0.82, respectively. It shows clearly that the proposed method outperforms the other two methods on DSC and HD metrics significantly. Figure 3 .B presents the ROC curves from fivefold cross-validation for our model other two methods, demonstrating an AUC of 0.95, 0.88 and 0.81, respectively. The mean contoured target volume derived from our method is 15.31 ± 8.65 cc, showing strong agreement with the GT mean volume of 15.05 ± 8.90 cc. Figure 3 .D illustrates that the average difference in contoured target volumes between the groups is 0.25 ± 1.39 cc. The correlation between bAVM size and the aforementioned geometrical indices are shown in Fig. 3 .E. The DSC is shown to be independent of bAVM size with a PCC of 0.02. The correlation between volume variation and HD demonstrates a moderate relationship (PCC = 0.53), albeit not exceptionally strong. These quantitative results all validate the high accuracy of contours delineated by our proposed segmentation method. Examples of manual and predicted contours are illustrated in Fig. 4 , with manual contours depicted in red, auto-segmented contours in blue, and discrepancies highlighted in yellow masks. 3.2 Effect of TBASE Assistance on the Readers’ Performance The use of TBASE assistance significantly enhanced contouring accuracy, resulting in a mean increase in the DSC of 0.10 ± 0.11 and HD of 0.78 ± 0.57. The mean DSC improved from 0.76 ± 0.07 in the unassisted mode to 0.86 ± 0.05 in the assisted mode ( P < 0.001). Corresponding values of mean HD reduced from 4.31 ± 0.68 to 3.35 ± 0.17 ( P < 0.001). The improvement in DSC and HD achieved with TBASE assistance for each reader is illustrated in Fig. 5 .A. Without assistance, non-specialist readers demonstrate notably lower contouring accuracy than specialist readers (DSC: 0.73 ± 0.13 vs. 0.83 ± 0.08, P < 0.05; HD: 4.63 ± 0.60 vs. 3.68 ± 0.33, P < 0.01). When using the TBASE, both groups performed similarly well (DSC, 0.85 ± 0.07 vs 0.88 ± 0.03, P = 0.286; HD, 3.62 ± 0.17 vs 3.34 ± 0.06, P = 0.114). Notably, non-specialist readers achieved greater improvements in contouring accuracy compared to specialists (DSC increase: 0.15 ± 0.11 vs 0.06 ± 0.09; P < 0.001; Fig. 5 .B). The mean contouring time per patient is significantly reduced with TBASE assistance compared to without (11 ± 7 min vs. 23 ± 9 min; P < 0.001), achieving a mean time saving of 52.15% ± 13.85%. Non-specialist readers demonstrated similar reductions in contouring time as specialists (50.68% ± 14.62% vs. 55.1% ± 11.98%; P = 0.217; Fig. 5 .C). 4. Discussion Segmentation of the eloquent bAVM generally challenging primarily due to the heterogeneity of the bAVM nidus, which includes abnormal vascular masses, brain parenchyma, cerebral spinal fluid, and both normal and embolized vessels, as well as the often-ill-defined boundaries of bAVM. In response, we proposed a two-stage bAVM segmentation ensemble TBASE utilizing the 2D U-Net detection model to localize the bAVM, and then, a 3D CMUnet-Cbam was developed to segment bAVM within identified location. Near-perfect DSC and HD are achieved for eloquent bAVM located adjacent to or within WM pathway, with average DSC and HD of 0.87 ± 0.03 and 3.51 ± 0.26. The DSC is considerably improved when compared to other networks such as Res-Net and U-Net, and HD is lowered. These results represent the highest reported segmentation performance for bAVM to date (as shown in Supplementary Table 1). In multi-reader study, we demonstrated that TBASE-assisted segmentation during eloquent bAVM contouring resulted in a significant improvement in contouring accuracy, with DSC and HD improvement of 0.86 and 3.35 mm. Additionally, there is a 52.15% reduction in workload for physicians with different-level of experience. Although there are emerging some reports on the segmentation of bAVM 17 , 18 , 25 , 26 , to our knowledge, this is the first research focus on accurate auto-segmentation of eloquent bAVM located adjacent to or within WM pathway using 3D-TOF-MRA and WM tractography. In our study, WM tractography in T1-weighted MRI image is employed in WM pathways visualization and segmentation to identify the anatomical relationship between WM pathways and the bAVM nidus, and then facilitate the protection for these critical areas of brain function. From the zoomed-in images in Fig. 4 , it is evident that the target area delineated by the TBASE model effectively avoided WM pathway in DTI images, closely matching the performance of manual delineation than other networks (as pointed out in Fig. 4 by yellow arrow). Meanwhile, the ResUnet tends to encounter challenges when accurately identifying embolized blood vessels, whereas the proposed method show no false bAVM boundary. (as pointed out in Fig. 4 by red arrow). Overall, our method successfully differentiated the bAVM from already embolized portion and surrounding normal vasculature. In this study, several strategies are implemented in the TBASE to improve the performances of our segmentation method. First, rather than performing segmentation directly on the whole input volume as traditional methods do, we employed a 2D detection network to generate a preliminary segmentation, identify the location and boundary of the bAVM nidus in the first step. And then, an optimized 3D segmentation network performed precise segmentation within these identified regions. By excluding irrelevant regions, we improve the accuracy of the target segmentation compared to conventional approaches. Second, compared with previous approaches using U-net for target segmentation, our CMUnet-Cbam architecture contains several unique features. On the one hand, CMUNeXt Block is introduced to thoroughly integrate distant spatial and location information, efficiently extracting global context information through large kernel and inverted bottleneck design 27 . On the other hand, we incorporate spatial and channel attention modules, along with a versatile "Attentional Block" to optimize performance while minimizing the number of parameters. Third, overfitting can be a critical issue, particularly given the relatively small dataset of 191 patients utilized in this study. Random deformation and rotation are applied to the training data to augment the number of training examples. This approach not only enhance the model’s ability to handle target deformation but also alleviate overfitting. Meanwhile, CMUNeXt Block is dedicated to reduce network parameters, further curtailing the risk of overfitting. Moreover, we adopt a five-fold cross-validation method to ensure that all the optimization strategies we applied are aimed at the training set. Table 2 DSC of TBASE-assisted and unassisted modes for evaluation of inter-reader variability among physicians with different-level of experience. TBASE-Assisted TBASE-Unassisted (mean ± std) (mean ± std) Non-Specialist NS-1 versus others 0.84 ± 0.04 0.77 ± 0.07 NS-2 versus others 0.87 ± 0.03 0.72 ± 0.12 NS-3 versus others 0.83 ± 0.06 0.78 ± 0.04 NS-4 versus others 0.87 ± 0.02 0.75 ± 0.06 Specialist S-1 versus S-2 0.89 0.82 The variability in target delineation remains a critical concern for the quality assurance of brain SRS, as highlighted in previous studies 28 . Our findings indicate that the use of TBASE assistance markedly enhanced inter-reader consistency across eloquent bAVM (Table 2 ), as exemplified by the case depicted in Fig. 6 .A. Notably, the improvement in inter-observer agreement is consistent across two specialists. These results suggest that collaborative intelligence could represent a significant advancement in ensuring high-quality outcomes in bAVM SRS. Meanwhile, the integration of TBASE significantly enhanced the accuracy of contouring, as evidenced by improved DSC and HD metrics. Clinicians assisted by TBASE exhibited notably greater precision in eloquent bAVM delineation compared to those working independently, particularly benefiting less-experienced clinicians. During unassisted contouring, the three clinicians are primarily focused on delineating the highlighted malformed vascular mass, which inadvertently led to an oversight of the normal cerebral artery, resulting in its erroneous inclusion within the target area (Fig. 6 .B). Arteries serve as critical conduits for cerebral perfusion, and excessive radiation exposure during treatment poses a significant risk of arterial wall injury, inflammation, stenosis, and may ultimately elevate the risk of cerebral ischemia or stroke. All clinicians successfully identified the cerebral artery and excluded it during contouring with the assistance of TBASE. The precise radiation dose gradient around the target implies that even minor improvements in delineation can significantly enhance bAVM occlusion while minimizing normal tissue toxicity. This study represents the first evaluation of AI assistance in managing bAVM, involving physicians across diverse experience levels. Compared to AI-assisted delineation of brain metastases, meningiomas, and acoustic neuromas 29 , delineating eloquent bAVM is considerably more challenging. This heightened complexity highlights the substantial potential for AI to significantly enhance clinical decision-making in this area. There are several limitations to this work: First, the experiment is conducted within only a single medical institution. Multicenter data collection would be useful for further validating and enhancing the robustness of the proposed auto-segment approach. Second, the segmented targets included some portion of other vasculature besides the nidus. Therefore, it is essential to quantify the percentages of nidus, brain tissue, and cerebrospinal fluid within the regions exposed to the prescribed isodose levels of radiation. Third, the sample size of multi-reader study may have been too small. Further research is required to elucidate the clinical significance of using collaborative intelligence in treating eloquent bAVM with SRS to facilitate prospective clinical validation. 5. Conclusion The reliable eloquent bAVM automated-segmentation method has been validated in clinical workflow. The TBASE assistance improved the accuracy and efficiency of the eloquent bAVM manual delineation. The integration of TBASE into current workflows has the potential to optimize SRS schemes and avoid neurological sequelae after SRS in considering WM pathways protection and thus improve patient care. Declarations Funding Statement This study was supported by the Science and Technology Innovation Action Plan of Shanghai Science and Technology Commission [Grant Nos. 24692107302 and 24692121500]. Dataset Availability The datasets described in this study are available from the corresponding author on reasonable request. Ethics approval and consent to participate This study was conducted in accordance with the Helsinki Declaration. Our Institutional Review Board of Huashan Hospital, Fudan University evaluated and approved the study design. The requirement for informed consent was waived by the ethics committe owing to the retrospective nature of the study. Consent for publication Not applicable. Interest Statement The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. References Lawton MT, Rutledge WC, Kim H, Stapf C, Whitehead KJ, Li DY, et al. Brain arteriovenous malformations. Nature reviews disease primers 2015;1:1-20. Fleetwood IG, Steinberg GK. Arteriovenous malformations. The Lancet 2002;359:863-873. Schauble B, Cascino G, Pollock B, Gorman D, Weigand S, Cohen-Gadol A, et al. Seizure outcomes after stereotactic radiosurgery for cerebral arteriovenous malformations. Neurology 2004;63:683-687. Davari M, Afzali M, Sabbagh-Bani-Azad M, Sharif Z, Hadizadeh F, Mirmoghtadaee P, et al. The Effectiveness of Selected Radiosurgery Devices on Cancer Tumors and Arteriovenous Malformations: A Systematic Review. Journal of Pharmacoeconomics and Pharmaceutical Management 2017;3:54-63. Luther N, Kondziolka D, Kano H, Mousavi SH, Flickinger JC, Lunsford LD. Motor function after stereotactic radiosurgery for brain metastases in the region of the motor cortex. Journal of neurosurgery 2013;119:683-688. Connor M, Karunamuni R, McDonald C, Seibert T, White N, Moiseenko V, et al. Regional susceptibility to dose-dependent white matter damage after brain radiotherapy. Radiotherapy and Oncology 2017;123:209-217. Hadjipanayis CG, Levy EI, Niranjan A, Firlik AD, Kondziolka D, Flickinger JC, et al. Stereotactic radiosurgery for motor cortex region arteriovenous malformations. Neurosurgery 2001;48:70-77. Conti A, Pontoriero A, Ricciardi GK, Granata F, Vinci S, Angileri FF, et al. Integration of functional neuroimaging in CyberKnife radiosurgery: feasibility and dosimetric results. Neurosurgical focus 2013;34:E5. Jeurissen B, Descoteaux M, Mori S, Leemans A. Diffusion MRI fiber tractography of the brain. NMR in Biomedicine 2019;32:e3785. Maruyama K, Kamada K, Ota T, Koga T, Itoh D, Ino K, et al. Tolerance of pyramidal tract to gamma knife radiosurgery based on diffusion-tensor tractography. International Journal of Radiation Oncology* Biology* Physics 2008;70:1330-1335. Krishnan AP, Asher IM, Davis D, Okunieff P, O'Dell WG. Evidence that MR diffusion tensor imaging (tractography) predicts the natural history of regional progression in patients irradiated conformally for primary brain tumors. International Journal of Radiation Oncology* Biology* Physics 2008;71:1553-1562. Sun L, Qu B, Wang J, Ju Z, Zhang Z, Cui Z, et al. Integration of functional MRI and white matter tractography in CyberKnife radiosurgery. Technology in cancer research & treatment 2017;16:850-856. Shi F, Hu W, Wu J, Han M, Wang J, Zhang W, et al. Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy. Nature communications 2022;13:6566. Yu N, Yu H, Li H, Ma N, Hu C, Wang J. A robust deep learning segmentation method for hematoma volumetric detection in intracerebral hemorrhage. Stroke 2022;53:167-176. Rauschecker AM, Gleason TJ, Nedelec P, Duong MT, Weiss DA, Calabrese E, et al. Interinstitutional portability of a deep learning brain MRI lesion segmentation algorithm. Radiology: Artificial Intelligence 2021;4:e200152. Tang H, Chen X, Liu Y, Lu Z, You J, Yang M, et al. Clinically applicable deep learning framework for organs at risk delineation in CT images. Nature Machine Intelligence 2019;1:480-491. Jiao Y, Zhang J-Z, Zhao Q, Liu J-Q, Wu Z-Z, Li Y, et al. Machine learning-enabled determination of diffuseness of brain arteriovenous malformations from magnetic resonance angiography. Translational Stroke Research 2022;13:939-948. Wang T, Lei Y, Tian S, Jiang X, Zhou J, Liu T, et al. Learning‐based automatic segmentation of arteriovenous malformations on contrast CT images in brain stereotactic radiosurgery. Medical physics 2019;46:3133-3141. Spetzler RF, Martin NA. A proposed grading system for arteriovenous malformations. Journal of neurosurgery 1986;65:476-483. Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, et al. N4ITK: improved N3 bias correction. IEEE transactions on medical imaging 2010;29:1310-1320. He K, Zhang X, Ren S, Sun J, editors. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition; 2016;pp.770-778. Ronneberger O, Fischer P, Brox T, editors. U-net: Convolutional networks for biomedical image segmentation. Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18(pp.234-241). Dice LR. Measures of the amount of ecologic association between species. Ecology 1945;26:297-302. Karimi D, Salcudean SE. Reducing the hausdorff distance in medical image segmentation with convolutional neural networks. IEEE Transactions on medical imaging 2019; 39:499-513. Hong JS, You WC, Sun MH, Pan HC, Lin YH, Lu YF, et al. Deep learning detection and segmentation of brain arteriovenous malformation on magnetic resonance angiography. Journal of Magnetic Resonance Imaging 2024;59:587-598. Wu S, Wu Y, Chang H, Su FT, Liao H, Tseng W, et al. Deep learning-based segmentation of various brain lesions for radiosurgery. Applied Sciences 2021;11(19):p9180. Tang F, Ding J, Wang L, Ning C, Zhou SK. Cmunext: An efficient medical image segmentation network based on large kernel and skip fusion. IEEE International Symposium on Biomedical Imaging (ISBI). 2024; (pp. 1-5). IEEE. Growcott S, Dembrey T, Patel R, Eaton D, Cameron A. Inter-observer variability in target volume delineations of benign and metastatic brain tumours for stereotactic radiosurgery: results of a national quality assurance programme. Clinical Oncology 2020;32:13-25. Lu S-L, Xiao F-R, Cheng JC-H, Yang W-C, Cheng Y-H, Chang Y-C, et al. Randomized multi-reader evaluation of automated detection and segmentation of brain tumors in stereotactic radiosurgery with deep neural networks. Neuro-oncology 2021;23:1560-1568. Additional Declarations No competing interests reported. Supplementary Files SupplementalMaterial.docx Cite Share Download PDF Status: Published Journal Publication published 02 Mar, 2026 Read the published version in Radiation Oncology → Version 1 posted Editorial decision: Revision requested 09 Dec, 2025 Reviews received at journal 08 Dec, 2025 Reviewers agreed at journal 26 Nov, 2025 Reviews received at journal 10 Jul, 2025 Reviewers agreed at journal 18 Jun, 2025 Reviewers invited by journal 22 May, 2025 Editor assigned by journal 15 May, 2025 Submission checks completed at journal 13 May, 2025 First submitted to journal 09 May, 2025 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-6632250","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":461513964,"identity":"426fceb3-4178-4d16-a3ab-60657d33d062","order_by":0,"name":"Mingzhu Li","email":"","orcid":"","institution":"The First Hospital of Hebei Medical University, Hebei Medical University","correspondingAuthor":false,"prefix":"","firstName":"Mingzhu","middleName":"","lastName":"Li","suffix":""},{"id":461513965,"identity":"0c36286e-951e-4fa4-b1ec-a018cdc2daa5","order_by":1,"name":"Huaguang Zhu","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Huaguang","middleName":"","lastName":"Zhu","suffix":""},{"id":461513966,"identity":"741fec4d-a2c8-4182-919a-531fe9aaa7ce","order_by":2,"name":"Yun Guan","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Yun","middleName":"","lastName":"Guan","suffix":""},{"id":461513967,"identity":"329905a7-2237-4e39-b863-16564fb4eaf0","order_by":3,"name":"Xiaoxia Liu","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Xiaoxia","middleName":"","lastName":"Liu","suffix":""},{"id":461513968,"identity":"9e0642fb-0f7e-4ff7-b595-15df1a6560c4","order_by":4,"name":"Xiaojia Gong","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Xiaojia","middleName":"","lastName":"Gong","suffix":""},{"id":461513969,"identity":"bd509b4b-b90f-4a1a-9541-2c0fbc5b75a2","order_by":5,"name":"Guanghai Mei","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Guanghai","middleName":"","lastName":"Mei","suffix":""},{"id":461513970,"identity":"29fc05c4-ebfb-4928-9f66-333296a47ddc","order_by":6,"name":"Tao Jin","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Tao","middleName":"","lastName":"Jin","suffix":""},{"id":461513971,"identity":"762e1f70-0402-4de4-b910-62de9441c3ca","order_by":7,"name":"Minghao Sun","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Minghao","middleName":"","lastName":"Sun","suffix":""},{"id":461513972,"identity":"c8587bac-5a8f-4c9a-a60b-c04c4d5c412f","order_by":8,"name":"Zhiyong Qin","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Zhiyong","middleName":"","lastName":"Qin","suffix":""},{"id":461513973,"identity":"2f841689-db30-4ef5-958a-99a96b231d01","order_by":9,"name":"Xing Di","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0ElEQVRIiWNgGAWjYDCCwwwMBxgYbKA8NuK1pJGi5QBMI9Fa+I7zGB4u+HU+z+D44QcMH8oOM/DPbsCvRfIwW8LhmX23iw3OpBkwzjh3mEHizgH8WgwOMx84zNtzO3HDgQQDZt62wwwGEgmEtDA2ALWcS9xw/vkH5r/EaQHawvPjQOKGGzkGzIzEaAH7hbchOXHmjTcFB3vOpfNI3CCghe/8GePPPH/sEvvOp2988KPMWo5/BgEtYMDYBqEPADEPEepB4A+R6kbBKBgFo2BkAgDCaEvG7vciEAAAAABJRU5ErkJggg==","orcid":"","institution":"Fudan University","correspondingAuthor":true,"prefix":"","firstName":"Xing","middleName":"","lastName":"Di","suffix":""}],"badges":[],"createdAt":"2025-05-10 03:53:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6632250/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6632250/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s13014-026-02811-2","type":"published","date":"2026-03-02T15:57:12+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":83480239,"identity":"8435a5f2-b7a0-4741-9d76-d10694060d09","added_by":"auto","created_at":"2025-05-27 06:32:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":259767,"visible":true,"origin":"","legend":"\u003cp\u003eThe schematic flow diagram of the proposed method. The upper part illustrates the training stage, while the lower part shows the testing stage for a new data.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6632250/v1/a96ef16c83c851a241264614.png"},{"id":83480242,"identity":"e4d02d03-83d1-41c0-ade9-1f6c544701e9","added_by":"auto","created_at":"2025-05-27 06:32:08","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":108786,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic diagrams of the multi-reader study design.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6632250/v1/950584ac6a46a00a98812beb.png"},{"id":83480246,"identity":"a5e23972-e2f4-4aed-a7a0-648890d83f59","added_by":"auto","created_at":"2025-05-27 06:32:09","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":10593931,"visible":true,"origin":"","legend":"\u003cp\u003eSegmentation performance of the TBASE on the test dataset. \u003cstrong\u003eA\u003c/strong\u003e: Distribution of DSC and HD for samples among all 38 patients in the test set in different networks. \u003cstrong\u003eB\u003c/strong\u003e: ROC curves from fivefold cross-validation for different networks. \u003cstrong\u003eC\u003c/strong\u003e: Linear relationship between the GT volumes and auto-segmented volumes. \u003cstrong\u003eD\u003c/strong\u003e: Difference plot of the two volume measurements. The dashed line represents the linear regression. \u003cstrong\u003eE\u003c/strong\u003e: Scatter\u003c/p\u003e\n\u003cp\u003eplots of the DSC (upper) and HD (lower).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6632250/v1/985754db6c876e1b7a62a6b9.png"},{"id":83480243,"identity":"63264041-4aa9-4ef0-b945-3b18c78de6f0","added_by":"auto","created_at":"2025-05-27 06:32:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":605359,"visible":true,"origin":"","legend":"\u003cp\u003eSegmentation results of TBASE and other methods for two tested cases in MRA and DTI views. The zoomed-in window location is indicated by red rectangle.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6632250/v1/79d470f644620fefe1607e14.png"},{"id":83480248,"identity":"d5bd3d34-c22e-44f4-8f7f-9faf28665b70","added_by":"auto","created_at":"2025-05-27 06:32:09","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":199404,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance comparison between unassisted and assisted readers. \u003cstrong\u003eA\u003c/strong\u003e: DSC (upper) and HD (lower) values for readers in unassisted (●) and assisted (▲) modes; * denotes a statistically significant difference. \u003cstrong\u003eB\u003c/strong\u003e: Improvement in DSC values and \u003cstrong\u003eC\u003c/strong\u003e: time savings for assisted readers across different experience levels.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6632250/v1/ba6e2823a868026f3cd6659a.png"},{"id":83480244,"identity":"3efd91d1-ecdf-4ef7-a04c-2c30b74f5f27","added_by":"auto","created_at":"2025-05-27 06:32:09","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":418896,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative case demonstrates increased inter-reader agreement and enhanced contouring accuracy with TBASE-assistance. Colored lines depict the contours drawn by four non-specialists, while the red line indicates the ground truth. \u003cstrong\u003eA:\u003c/strong\u003eCollaboration with TBASE markedly increased inter-reader consistency among non-specialists. \u003cstrong\u003eB:\u003c/strong\u003e TBASE assistance led to substantial improvements in contouring accuracy.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6632250/v1/1c0a2dc01bcf69db29ab37a6.png"},{"id":104250700,"identity":"28bdea09-46d6-44d2-b6b6-0e5d7114e393","added_by":"auto","created_at":"2026-03-09 16:05:58","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":11428166,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6632250/v1/dd676ae2-06d2-46c3-9cd5-fdfd75453062.pdf"},{"id":83480245,"identity":"b11f343a-93d9-4564-b0c4-8f7f054cae4a","added_by":"auto","created_at":"2025-05-27 06:32:09","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":2329391,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-6632250/v1/30f178d727bdbe65c01f42db.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multi-Reader Evaluation of Deep Learning-Based Auto-Segmentation of Eloquent Brain Arteriovenous Malformation on MRA and White Matter Tractography in Stereotactic Radiosurgery","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBrain arteriovenous malformation is a vascular anomaly typically characterized by abnormal connections between arteries and veins, potentially leading to disrupted blood flow and increased risk of bleeding \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Stereotactic radiosurgery, including technologies like CyberKnife, has been widely recognized as an effective therapeutic approach for managing bAVM \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Due to the therapeutic benefits of delivering maximal radiation dose, a crucial aspect of radiosurgery is determining the optimal radiation margin while safeguarding language, visual, and motor functions \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Hence, accurately localizing eloquent cortical regions and WM pathways within the bAVM margin is imperative for avoiding neurological sequelae \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Recently, WM tractography has been incorporated into the treatment planning process for SRS \u003csup\u003e\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. This integration aided in identifying the anatomical relationship between WM fibers and nidus, thereby facilitating the optimization of treatment plan design \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Accurate identification and delineation of the bAVM nidus are essential stages in SRS treatment planning. Presently, targets are manually identified by neuroradiologists and contoured by radiation oncologists, utilizing multi-modal and WM Tractography images while referring to digital subtraction angiography images. However, this entire process is often time-consuming, labor-intensive, and susceptible to inter-observer variability, especially when dealing with locations adjacent to WM pathways.\u003c/p\u003e \u003cp\u003eWith DL algorithms represent the forefront of medical image analysis \u003csup\u003e\u003cspan additionalcitationids=\"CR14 CR15\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, auto-segmentation is devised to address the limitations of bAVM\u0026rsquo;s manual contours by delivering faster and observer-independent results. A U-Net model was proposed by Jiao et al. to detect and quantify the diffuseness of bAVM nidus in MRA image \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Wang et al. developed a deeply supervised V-Net to delineate the bAVM nidus from CT images and evaluated target dose coverage changes on original plan \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Both studies have validated the reliability of using DL for the direct segmentation of bAVM nidus on medical images. However, there are still a lot of limitations. These methodologies predominantly rely on single-modal images, which may not adequately capture the complex morphology and characteristics of bAVM. Furthermore, lesions located near WM pathways often received radiation doses exceeding safe thresholds due to difficulties in their precise identification on CT or conventional MRI scans. Lastly, despite the promising standalone performance of DL in preclinical evaluations of bAVM SRS, there has been a significant lack of clinical trials to evaluate how these results translate into practical clinical settings. To fully assess the clinical potential of DL in managing bAVM, it is crucial to incorporate the system into clinical practice and evaluate its effect on practitioners.\u003c/p\u003e \u003cp\u003eGiven the limitations mentioned above, we have developed an innovative two-stage DL ensemble (TBASE), which integrated 2D U-Net detection-aided and 3D CMUnet-Cbam for automatic bAVM segmentation. This approach leverage multi-modal imaging alongside WM tractography, marking the first utilization of WM tractography in the auto-segmentation of bAVM. It aims to enhance the precision of target data extraction, thus ensuring the preservation of WM pathways. The results of bAVM auto-segmentations are evaluated for standard geometric indices including DSC and HD. Subsequently, we conducted a cross-modal, multi-reader study involving physicians across various SRS experience levels and specialties, utilizing TBASE for target contouring of eloquent bAVM. We assessed inter-reader variability, accuracy, and the efficiency of assisted contouring for eloquent bAVM SRS.\u003c/p\u003e"},{"header":"2. Methods and Materials","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Clinical dataset\u003c/h2\u003e \u003cp\u003eA retrospective analysis is conducted on 191 patients with bAVM located either adjacent to or within WM pathways, all of whom underwent CyberKnife radiosurgery at Huashan Hospital, Fudan University, from 2016 and 2023. Our retrospective study was approved by the Ethical Review Committee of the Huashan Hospital, Fudan University and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived owing to the retrospective nature of the study. Demographic and clinical characteristics, including age, sex, lesion size, affected hemisphere, specific lesion location, and Spetzler-Martin grading \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e are documented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Statistical analysis revealed no significant differences in these baseline characteristics between the training and testing sets utilized in the segmentation model. All manual contours are rigorously identified by neuroradiologists and contoured by radiation oncologists to ensure quality control. Image acquisition information is shown in Supplementary Appendix A.\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\u003eDemographic and clinical characteristics of patients for bAVM segmentation.\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=\"left\" 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\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTraining Set\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;153)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTest Set\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;38)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, median [IQR]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e26 [17\u0026ndash;62]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e25 [19\u0026ndash;48]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eGender, number (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e66 (43.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18 (47.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e87 (56.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e20 (52.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSize (cc), median [IQR]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15.4 [1.3\u0026ndash;55.9]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e15.1 [1.5\u0026ndash;48.6]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eLesion-affected hemisphere.\u003c/p\u003e \u003cp\u003enumber (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLeft\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e74 (48.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18 (47.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e79 (51.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e20 (52.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eLesion location, number (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFrontal lobe\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e51 (33.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13 (34.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eParietal lobe\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e29 (19%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12 (31.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBasal ganglia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e73 (47.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13 (34.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eS-M grading, number (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e61 (39.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14 (36.8%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e84 (54.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23 (60.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (5.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (2.6%)\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 \u003cem\u003eIQR\u003c/em\u003e interquartile range, \u003cem\u003eS-M grading\u003c/em\u003e Spetzler-Martin grading.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Data preprocessing\u003c/h2\u003e \u003cp\u003eTo construct TBASE, the 153 patients are divided into training and validation group with a ratio of approximately 80%: 20% (122: 31), and the other 38 patients are reserved to test performance according to different locations (frontal lobe, parietal lobe and basal ganglia). Image registration of each patient\u0026rsquo;s WM tractography to their corresponding MRA image is accomplished using a proprietary MIM workflow \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. To enhance the robustness of the model, data augmentation such as rotation, flipping, and zooming are employed, facilitating the model's ability to learn transformation-invariant features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Segmentation\u003c/h2\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.3.1 Overview\u003c/h2\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eA two-stage deep learning ensemble TBASE is trained to segment bAVM nidus, which consisting of a 2D U-Net detection model to localize the bAVM ROI and 3D CMUnet-Cbam to segment bAVM within identified location. The whole processes for training and validation phase are illustrated in upper part of Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. In the first stage of training phase, the multimodal images of each training patient are served as input images and concatenated as multi-channel inputs, with the manual contours serving as the target for learning. The binary mask of bAVM is coarsely generated in 2D U-Net and then a 128\u0026times;128\u0026times;128 pixels ROI bounding box centered at the mass center of mask is created to localize the bAVM nidus. The coordinates of these cropped ROIs are stored and subsequently utilized to map the predicted masks back to the original images. In the second stage, the multi-modal images are cropped to the defined ROI, and these images are then input into the 3D CMUnet-Cbam architecture, which provided accurate and efficient bAVM segmentation within the ROI. In the testing phase, for a new multi-modal image dataset, the localization of the bAVM ROI and subsequent segmentation within that ROI are achieved following the same process as in the training phase. The final segmented bAVM contour is then reconstructed by positioning the segmented bAVM ROI within the full 3D image volume. The 2D U-Net and 3D CMUnet-Cbam models are implemented using Python 3.8 using Pytorch 1.12, enhanced by CUDA 12.2, on six NVIDIA RTX 4080Ti GPUs. The models underwent training and testing across 300 epochs, employing the Adam optimizer, which incorporated a momentum term of 0.5 to enhance convergence. The initial learning rate is set at 0.0002 and systematically reduced by half whenever the error rates on the validation set plateaued, ensuring efficient and effective learning progression.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.3.2 Network architecture\u003c/h2\u003e \u003cp\u003eThe architecture of our proposed CMUnet-Cbam is detailed in bottom part of FIG.1, and it is divided into two stages: encoding and decoding. In the encoding stage, which includes five levels organized from top to bottom, the input image slices which cropped in the detection phase initially pass through a combination of conventional convolution and attention blocks, with max-pooling employed to reduce the size of the feature maps. And the following four levels utilize the CMUNeXt block paired with an attentional block. To maintain the multi-level features effectively, concatenation is employed to enhance the volume of feature maps from both the encoding and decoding stages. The final transformation involves funneling these maps through a convolution layer to decrease their dimensionality to two channels. This is immediately followed by a \u0026ldquo;tanh\u0026rdquo; activation layer that serves to polarize the feature maps, thus distinguishing the posterior probabilities of the bAVM nidus from normal tissue. The network specifications are described in detail in Supplementary Appendix B. To validate the effectiveness of the proposed network, we conducted comparative experiments with Res-Net and U-Net \u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Multi-reader study design\u003c/h2\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe study employs a cross-modal, multi-reader design as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. A separate dataset comprising of ten bAVM patients are randomly selected from test dataset for the reader study. All cases are independent of the development dataset used for model training. Two specialist and four non-specialist SRS physicians contoured the selected dataset cases, with or without TBASE assistance. To minimize potential bias from previous contouring, participants are randomly assigned to either the assisted-first or unassisted-first segmentation mode. The two contour sessions are separated by an interval of at least 6-weeks. During the second session, physicians are blinded to both their initial segmentations and those from other participants. For this separate dataset, the GT is independently delineated by two SRS experts. Three-dimensional reconstructions of their contours are simultaneously visualized and compared, allowing for direct assessment of inter-expert variability. Any discrepancies observed are resolved through a consensus process.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Evolution metrics and Statistical Analysis\u003c/h2\u003e \u003cp\u003eQuantitative comparisons between the segmentation results from our proposed model and manual contouring are conducted for each patient in the validation dataset and multi-reader study cases, employing metrics such as DSC and HD \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. The metrics employed in this study are defined as follows:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:DSC(A,B)=\\frac{2|A\\cap\\:B|}{\\left|A\\right|+\\left|B\\right|}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:HD(A,B)=\\text{max}\\left(\\text{h}\\left(A,B\\right),\\text{h}\\left(B,A\\right)\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere A and B represent auto-segment and manual contours, respectively.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:\\text{h}\\text{(}A,B\\text{)}\\text{=}{\\text{max}}_{\\text{a}\\in\\text{A}}{\\text{min}}_{\\text{b}\\in\\text{B}}\\left|\\right|\\text{a}\\text{-}\\text{b||}\\)\u003c/span\u003e \u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{h}\\text{(}B,A\\text{)}\\text{=}{\\text{max}}_{\\text{b}\\in\\text{B}}{\\text{min}}_{\\text{a}\\in\\text{A}}\\left|\\right|\\text{b}\\text{-}\\text{a}\\text{||}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left|\\right|\\text{.}\\text{||}\\)\u003c/span\u003e\u003c/span\u003e denotes the Euclidean distance.\u003c/p\u003e \u003cp\u003eWe conducted Wilcoxon paired-signed tests to compare the DSC, HD and time consumed between two contouring modes for bAVM detection. The statistical significance is determined with a two-tailed \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05. Meanwhile, the Pearson correlation coefficient (PCC) is used to evaluate the influence of bAVM size on the DSC and HD metrics.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 TBASE Performance\u003c/h2\u003e\n \u003cp\u003eThe overall geometric parameters for all 38 patients in test dataset and model evaluation are numerical summarized in FIG. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eAs shown in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.A, the mean values and standard deviations for DSC and HD are 0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03 and 3.51\u0026thinsp;\u0026plusmn;\u0026thinsp;0.26, respectively, while Res-Net and U-Net results are 0.75\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12 and 4.14\u0026thinsp;\u0026plusmn;\u0026thinsp;0.99, 0.77\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09 and 3.94\u0026thinsp;\u0026plusmn;\u0026thinsp;0.82, respectively. It shows clearly that the proposed method outperforms the other two methods on DSC and HD metrics significantly. Figure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.B presents the ROC curves from fivefold cross-validation for our model other two methods, demonstrating an AUC of 0.95, 0.88 and 0.81, respectively. The mean contoured target volume derived from our method is 15.31\u0026thinsp;\u0026plusmn;\u0026thinsp;8.65 cc, showing strong agreement with the GT mean volume of 15.05\u0026thinsp;\u0026plusmn;\u0026thinsp;8.90 cc. Figure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.D illustrates that the average difference in contoured target volumes between the groups is 0.25\u0026thinsp;\u0026plusmn;\u0026thinsp;1.39 cc. The correlation between bAVM size and the aforementioned geometrical indices are shown in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.E. The DSC is shown to be independent of bAVM size with a PCC of 0.02. The correlation between volume variation and HD demonstrates a moderate relationship (PCC\u0026thinsp;=\u0026thinsp;0.53), albeit not exceptionally strong. These quantitative results all validate the high accuracy of contours delineated by our proposed segmentation method. Examples of manual and predicted contours are illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e, with manual contours depicted in red, auto-segmented contours in blue, and discrepancies highlighted in yellow masks.\u003c/p\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Effect of TBASE Assistance on the Readers\u0026rsquo; Performance\u003c/h2\u003e\n \u003cp\u003eThe use of TBASE assistance significantly enhanced contouring accuracy, resulting in a mean increase in the DSC of 0.10\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11 and HD of 0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.57. The mean DSC improved from 0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07 in the unassisted mode to 0.86\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05 in the assisted mode (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Corresponding values of mean HD reduced from 4.31\u0026thinsp;\u0026plusmn;\u0026thinsp;0.68 to 3.35\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17 (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). The improvement in DSC and HD achieved with TBASE assistance for each reader is illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e.A.\u003c/p\u003e\n \u003cp\u003eWithout assistance, non-specialist readers demonstrate notably lower contouring accuracy than specialist readers (DSC: 0.73\u0026thinsp;\u0026plusmn;\u0026thinsp;0.13 vs. 0.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05; HD: 4.63\u0026thinsp;\u0026plusmn;\u0026thinsp;0.60 vs. 3.68\u0026thinsp;\u0026plusmn;\u0026thinsp;0.33, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01). When using the TBASE, both groups performed similarly well (DSC, 0.85\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07 vs 0.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.286; HD, 3.62\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17 vs 3.34\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.114). Notably, non-specialist readers achieved greater improvements in contouring accuracy compared to specialists (DSC increase: 0.15\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11 vs 0.06\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09; \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001; Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e.B). The mean contouring time per patient is significantly reduced with TBASE assistance compared to without (11\u0026thinsp;\u0026plusmn;\u0026thinsp;7 min vs. 23\u0026thinsp;\u0026plusmn;\u0026thinsp;9 min; \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), achieving a mean time saving of 52.15% \u0026plusmn; 13.85%. Non-specialist readers demonstrated similar reductions in contouring time as specialists (50.68% \u0026plusmn; 14.62% vs. 55.1% \u0026plusmn; 11.98%; \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.217; Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e.C).\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eSegmentation of the eloquent bAVM generally challenging primarily due to the heterogeneity of the bAVM nidus, which includes abnormal vascular masses, brain parenchyma, cerebral spinal fluid, and both normal and embolized vessels, as well as the often-ill-defined boundaries of bAVM. In response, we proposed a two-stage bAVM segmentation ensemble TBASE utilizing the 2D U-Net detection model to localize the bAVM, and then, a 3D CMUnet-Cbam was developed to segment bAVM within identified location. Near-perfect DSC and HD are achieved for eloquent bAVM located adjacent to or within WM pathway, with average DSC and HD of 0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03 and 3.51\u0026thinsp;\u0026plusmn;\u0026thinsp;0.26. The DSC is considerably improved when compared to other networks such as Res-Net and U-Net, and HD is lowered. These results represent the highest reported segmentation performance for bAVM to date (as shown in Supplementary Table\u0026nbsp;1). In multi-reader study, we demonstrated that TBASE-assisted segmentation during eloquent bAVM contouring resulted in a significant improvement in contouring accuracy, with DSC and HD improvement of 0.86 and 3.35 mm. Additionally, there is a 52.15% reduction in workload for physicians with different-level of experience.\u003c/p\u003e \u003cp\u003eAlthough there are emerging some reports on the segmentation of bAVM \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e,\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e, to our knowledge, this is the first research focus on accurate auto-segmentation of eloquent bAVM located adjacent to or within WM pathway using 3D-TOF-MRA and WM tractography. In our study, WM tractography in T1-weighted MRI image is employed in WM pathways visualization and segmentation to identify the anatomical relationship between WM pathways and the bAVM nidus, and then facilitate the protection for these critical areas of brain function. From the zoomed-in images in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, it is evident that the target area delineated by the TBASE model effectively avoided WM pathway in DTI images, closely matching the performance of manual delineation than other networks (as pointed out in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e by yellow arrow). Meanwhile, the ResUnet tends to encounter challenges when accurately identifying embolized blood vessels, whereas the proposed method show no false bAVM boundary. (as pointed out in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e by red arrow). Overall, our method successfully differentiated the bAVM from already embolized portion and surrounding normal vasculature.\u003c/p\u003e \u003cp\u003eIn this study, several strategies are implemented in the TBASE to improve the performances of our segmentation method. First, rather than performing segmentation directly on the whole input volume as traditional methods do, we employed a 2D detection network to generate a preliminary segmentation, identify the location and boundary of the bAVM nidus in the first step. And then, an optimized 3D segmentation network performed precise segmentation within these identified regions. By excluding irrelevant regions, we improve the accuracy of the target segmentation compared to conventional approaches. Second, compared with previous approaches using U-net for target segmentation, our CMUnet-Cbam architecture contains several unique features. On the one hand, CMUNeXt Block is introduced to thoroughly integrate distant spatial and location information, efficiently extracting global context information through large kernel and inverted bottleneck design \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. On the other hand, we incorporate spatial and channel attention modules, along with a versatile \"Attentional Block\" to optimize performance while minimizing the number of parameters. Third, overfitting can be a critical issue, particularly given the relatively small dataset of 191 patients utilized in this study. Random deformation and rotation are applied to the training data to augment the number of training examples. This approach not only enhance the model\u0026rsquo;s ability to handle target deformation but also alleviate overfitting. Meanwhile, CMUNeXt Block is dedicated to reduce network parameters, further curtailing the risk of overfitting. Moreover, we adopt a five-fold cross-validation method to ensure that all the optimization strategies we applied are aimed at the training set.\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\u003eDSC of TBASE-assisted and unassisted modes for evaluation of inter-reader variability among physicians with different-level of experience.\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\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTBASE-Assisted\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTBASE-Unassisted\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e(mean\u0026thinsp;\u0026plusmn;\u0026thinsp;std)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNon-Specialist\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNS-1\u003c/p\u003e \u003cp\u003eversus others\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.84\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.77\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNS-2\u003c/p\u003e \u003cp\u003eversus others\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.72\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNS-3\u003c/p\u003e \u003cp\u003eversus others\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNS-4\u003c/p\u003e \u003cp\u003eversus others\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.75\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSpecialist\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eS-1\u003c/p\u003e \u003cp\u003eversus S-2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.82\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\u003eThe variability in target delineation remains a critical concern for the quality assurance of brain SRS, as highlighted in previous studies \u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. Our findings indicate that the use of TBASE assistance markedly enhanced inter-reader consistency across eloquent bAVM (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), as exemplified by the case depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.A. Notably, the improvement in inter-observer agreement is consistent across two specialists. These results suggest that collaborative intelligence could represent a significant advancement in ensuring high-quality outcomes in bAVM SRS. Meanwhile, the integration of TBASE significantly enhanced the accuracy of contouring, as evidenced by improved DSC and HD metrics. Clinicians assisted by TBASE exhibited notably greater precision in eloquent bAVM delineation compared to those working independently, particularly benefiting less-experienced clinicians. During unassisted contouring, the three clinicians are primarily focused on delineating the highlighted malformed vascular mass, which inadvertently led to an oversight of the normal cerebral artery, resulting in its erroneous inclusion within the target area (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.B). Arteries serve as critical conduits for cerebral perfusion, and excessive radiation exposure during treatment poses a significant risk of arterial wall injury, inflammation, stenosis, and may ultimately elevate the risk of cerebral ischemia or stroke. All clinicians successfully identified the cerebral artery and excluded it during contouring with the assistance of TBASE. The precise radiation dose gradient around the target implies that even minor improvements in delineation can significantly enhance bAVM occlusion while minimizing normal tissue toxicity. This study represents the first evaluation of AI assistance in managing bAVM, involving physicians across diverse experience levels. Compared to AI-assisted delineation of brain metastases, meningiomas, and acoustic neuromas \u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e, delineating eloquent bAVM is considerably more challenging. This heightened complexity highlights the substantial potential for AI to significantly enhance clinical decision-making in this area.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThere are several limitations to this work: First, the experiment is conducted within only a single medical institution. Multicenter data collection would be useful for further validating and enhancing the robustness of the proposed auto-segment approach. Second, the segmented targets included some portion of other vasculature besides the nidus. Therefore, it is essential to quantify the percentages of nidus, brain tissue, and cerebrospinal fluid within the regions exposed to the prescribed isodose levels of radiation. Third, the sample size of multi-reader study may have been too small. Further research is required to elucidate the clinical significance of using collaborative intelligence in treating eloquent bAVM with SRS to facilitate prospective clinical validation.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThe reliable eloquent bAVM automated-segmentation method has been validated in clinical workflow. The TBASE assistance improved the accuracy and efficiency of the eloquent bAVM manual delineation. The integration of TBASE into current workflows has the potential to optimize SRS schemes and avoid neurological sequelae after SRS in considering WM pathways protection and thus improve patient care.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by the Science and Technology Innovation Action Plan of Shanghai Science and Technology Commission [Grant Nos. 24692107302 and 24692121500].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDataset Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets described in this study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was conducted in accordance with the Helsinki Declaration. Our Institutional Review Board of Huashan Hospital, Fudan University evaluated and approved the study design. The requirement for informed consent was waived by the ethics committe owing to the retrospective nature of the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInterest Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eLawton MT, Rutledge WC, Kim H, Stapf C, Whitehead KJ, Li DY, et al. Brain arteriovenous malformations. Nature reviews disease primers 2015;1:1-20.\u003c/li\u003e\n \u003cli\u003eFleetwood IG, Steinberg GK. Arteriovenous malformations. The Lancet 2002;359:863-873.\u003c/li\u003e\n \u003cli\u003eSchauble B, Cascino G, Pollock B, Gorman D, Weigand S, Cohen-Gadol A, et al. Seizure outcomes after stereotactic radiosurgery for cerebral arteriovenous malformations. Neurology 2004;63:683-687.\u003c/li\u003e\n \u003cli\u003eDavari M, Afzali M, Sabbagh-Bani-Azad M, Sharif Z, Hadizadeh F, Mirmoghtadaee P, et al. The Effectiveness of Selected Radiosurgery Devices on Cancer Tumors and Arteriovenous Malformations: A Systematic Review. Journal of Pharmacoeconomics and Pharmaceutical Management 2017;3:54-63.\u003c/li\u003e\n \u003cli\u003eLuther N, Kondziolka D, Kano H, Mousavi SH, Flickinger JC, Lunsford LD. Motor function after stereotactic radiosurgery for brain metastases in the region of the motor cortex. Journal of neurosurgery 2013;119:683-688.\u003c/li\u003e\n \u003cli\u003eConnor M, Karunamuni R, McDonald C, Seibert T, White N, Moiseenko V, et al. Regional susceptibility to dose-dependent white matter damage after brain radiotherapy. Radiotherapy and Oncology 2017;123:209-217.\u003c/li\u003e\n \u003cli\u003eHadjipanayis CG, Levy EI, Niranjan A, Firlik AD, Kondziolka D, Flickinger JC, et al. Stereotactic radiosurgery for motor cortex region arteriovenous malformations. Neurosurgery 2001;48:70-77.\u003c/li\u003e\n \u003cli\u003eConti A, Pontoriero A, Ricciardi GK, Granata F, Vinci S, Angileri FF, et al. Integration of functional neuroimaging in CyberKnife radiosurgery: feasibility and dosimetric results. Neurosurgical focus 2013;34:E5.\u003c/li\u003e\n \u003cli\u003eJeurissen B, Descoteaux M, Mori S, Leemans A. Diffusion MRI fiber tractography of the brain. NMR in Biomedicine 2019;32:e3785.\u003c/li\u003e\n \u003cli\u003eMaruyama K, Kamada K, Ota T, Koga T, Itoh D, Ino K, et al. Tolerance of pyramidal tract to gamma knife radiosurgery based on diffusion-tensor tractography. International Journal of Radiation Oncology* Biology* Physics 2008;70:1330-1335.\u003c/li\u003e\n \u003cli\u003eKrishnan AP, Asher IM, Davis D, Okunieff P, O\u0026apos;Dell WG. Evidence that MR diffusion tensor imaging (tractography) predicts the natural history of regional progression in patients irradiated conformally for primary brain tumors. International Journal of Radiation Oncology* Biology* Physics 2008;71:1553-1562.\u003c/li\u003e\n \u003cli\u003eSun L, Qu B, Wang J, Ju Z, Zhang Z, Cui Z, et al. Integration of functional MRI and white matter tractography in CyberKnife radiosurgery. Technology in cancer research \u0026amp; treatment 2017;16:850-856.\u003c/li\u003e\n \u003cli\u003eShi F, Hu W, Wu J, Han M, Wang J, Zhang W, et al. Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy. Nature communications 2022;13:6566.\u003c/li\u003e\n \u003cli\u003eYu N, Yu H, Li H, Ma N, Hu C, Wang J. A robust deep learning segmentation method for hematoma volumetric detection in intracerebral hemorrhage. Stroke 2022;53:167-176.\u003c/li\u003e\n \u003cli\u003eRauschecker AM, Gleason TJ, Nedelec P, Duong MT, Weiss DA, Calabrese E, et al. Interinstitutional portability of a deep learning brain MRI lesion segmentation algorithm. Radiology: Artificial Intelligence 2021;4:e200152.\u003c/li\u003e\n \u003cli\u003eTang H, Chen X, Liu Y, Lu Z, You J, Yang M, et al. Clinically applicable deep learning framework for organs at risk delineation in CT images. Nature Machine Intelligence 2019;1:480-491.\u003c/li\u003e\n \u003cli\u003eJiao Y, Zhang J-Z, Zhao Q, Liu J-Q, Wu Z-Z, Li Y, et al. Machine learning-enabled determination of diffuseness of brain arteriovenous malformations from magnetic resonance angiography. Translational Stroke Research 2022;13:939-948.\u003c/li\u003e\n \u003cli\u003eWang T, Lei Y, Tian S, Jiang X, Zhou J, Liu T, et al. Learning‐based automatic segmentation of arteriovenous malformations on contrast CT images in brain stereotactic radiosurgery. Medical physics 2019;46:3133-3141.\u003c/li\u003e\n \u003cli\u003eSpetzler RF, Martin NA. A proposed grading system for arteriovenous malformations. Journal of neurosurgery 1986;65:476-483.\u003c/li\u003e\n \u003cli\u003eTustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, et al. N4ITK: improved N3 bias correction. IEEE transactions on medical imaging 2010;29:1310-1320.\u003c/li\u003e\n \u003cli\u003eHe K, Zhang X, Ren S, Sun J, editors. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition; 2016;pp.770-778.\u003c/li\u003e\n \u003cli\u003eRonneberger O, Fischer P, Brox T, editors. U-net: Convolutional networks for biomedical image segmentation. Medical image computing and computer-assisted intervention\u0026ndash;MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18(pp.234-241).\u003c/li\u003e\n \u003cli\u003eDice LR. Measures of the amount of ecologic association between species. Ecology 1945;26:297-302.\u003c/li\u003e\n \u003cli\u003eKarimi D, Salcudean SE. Reducing the hausdorff distance in medical image segmentation with convolutional neural networks. IEEE Transactions on medical imaging 2019; 39:499-513.\u003c/li\u003e\n \u003cli\u003eHong JS, You WC, Sun MH, Pan HC, Lin YH, Lu YF, et al. Deep learning detection and segmentation of brain arteriovenous malformation on magnetic resonance angiography. Journal of Magnetic Resonance Imaging 2024;59:587-598.\u003c/li\u003e\n \u003cli\u003eWu S, Wu Y, Chang H, Su FT, Liao H, Tseng W, et al. Deep learning-based segmentation of various brain lesions for radiosurgery. Applied Sciences 2021;11(19):p9180.\u003c/li\u003e\n \u003cli\u003eTang F, Ding J, Wang L, Ning C, Zhou SK. Cmunext: An efficient medical image segmentation network based on large kernel and skip fusion. IEEE International Symposium on Biomedical Imaging (ISBI). 2024; (pp. 1-5). IEEE.\u003c/li\u003e\n \u003cli\u003eGrowcott S, Dembrey T, Patel R, Eaton D, Cameron A. Inter-observer variability in target volume delineations of benign and metastatic brain tumours for stereotactic radiosurgery: results of a national quality assurance programme. Clinical Oncology 2020;32:13-25.\u003c/li\u003e\n \u003cli\u003eLu S-L, Xiao F-R, Cheng JC-H, Yang W-C, Cheng Y-H, Chang Y-C, et al. Randomized multi-reader evaluation of automated detection and segmentation of brain tumors in stereotactic radiosurgery with deep neural networks. Neuro-oncology 2021;23:1560-1568.\u003c/li\u003e\n\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":"radiation-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"raon","sideBox":"Learn more about [Radiation Oncology](http://ro-journal.biomedcentral.com/)","snPcode":"13014","submissionUrl":"https://submission.nature.com/new-submission/13014/3","title":"Radiation Oncology","twitterHandle":"@OncoBioMed","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Brain arteriovenous malformation, White matter tractography, Stereotactic radiosurgery, Deep learning, Automatic segmentation, Randomized multi-reader","lastPublishedDoi":"10.21203/rs.3.rs-6632250/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6632250/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eObjective: \u003c/strong\u003eTo minimize the radiation injury for white matter (WM) pathways during brain arteriovenous malformation (bAVM) stereotactic radiosurgery (SRS), the WM tractography is integrated into treatment planning to identify WM pathways and restrict receiving dose. Manual segmentation of eloquent bAVM adjacent to WM pathways is time-consuming and prone to substantial inter-practitioner variability due to intricate entanglement within eloquent brain areas. The objective of this study is to develop and evaluate a deep learning (DL) system for the segmentation of eloquent bAVM in a clinical setting.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eA total of 191 eloquent bAVM patients who underwent WM tractography and 3D time-of-flight magnetic resonance angiography (TOF-MRA) images were enrolled. 153 patients were used to construct a two-stage DL bAVM segmentation ensemble (TBASE) consisting of 2D detection and 3D segmentation models to segment the bAVM, the other 38 to test performance. Comparative experiments with Res-Net and U-Net were conducted to validate the effectiveness of the proposed network. A randomized multi-reader evaluation was then conducted to assess the impact of TBASE assistance for bAVM segmentation using ten algorithm-unseen cases. Six medical professionals contoured the same series of cases in both assisted and unassisted modes, with a 6-week memory washout period between each session. The aided and unaided Dice Similarity Coefficients (DSC), Hausdorff Distance (HD), along with contouring times were compared.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e The mean values and standard deviations for DSC and HD of TBASE are 0.87 ± 0.03 and 3.51 ± 0.26, respectively, while Res-Net and U-Net results are 0.75 ± 0.12 and 4.14 ± 0.99, 0.77 ± 0.09 and 3.94 ± 0.82, respectively. The average volume difference across all patients in test dataset is 0.25 ± 1.39 cc, with no statistically significant variation observed. With TBASE assistance, the mean DSC of readers improved from 0.76 ± 0.07 to 0.86 ± 0.05 (\u003cem\u003eP\u003c/em\u003e\u0026lt; 0.001), with corresponding values of mean HD reducing from 4.31 ± 0.68 to 3.35 ± 0.17 (\u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001) and a mean time saving of 52.15% ± 13.85% per patient. Less-experienced readers achieved greater improvements in contouring accuracy compared to specialists (DSC increase: 0.15 ± 0.11 vs 0.06 ± 0.09; \u003cem\u003eP\u003c/em\u003e\u0026lt; 0.001), while demonstrating similar reductions in contouring time as specialists (50.68% ± 14.62% vs. 55.1% ± 11.98%; \u003cem\u003eP\u003c/em\u003e = 0.217)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003e The reliable eloquent bAVM automated-segmentation method has been validated in clinical workflow. The TBASE assistance improved the accuracy and efficiency of the eloquent bAVM manual delineation and thus avoid neurological sequelae after SRS in considering WM pathways protection.\u003c/p\u003e","manuscriptTitle":"Multi-Reader Evaluation of Deep Learning-Based Auto-Segmentation of Eloquent Brain Arteriovenous Malformation on MRA and White Matter Tractography in Stereotactic Radiosurgery","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-27 06:32:03","doi":"10.21203/rs.3.rs-6632250/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-12-09T19:05:00+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-08T13:17:37+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"274275945514633366002594813811694480357","date":"2025-11-26T23:41:45+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-10T09:24:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"122875696409190553306403455214057685449","date":"2025-06-19T02:33:07+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-22T11:41:33+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-16T00:08:19+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-05-14T02:01:45+00:00","index":"","fulltext":""},{"type":"submitted","content":"Radiation Oncology","date":"2025-05-10T03:47:55+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"radiation-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"raon","sideBox":"Learn more about [Radiation Oncology](http://ro-journal.biomedcentral.com/)","snPcode":"13014","submissionUrl":"https://submission.nature.com/new-submission/13014/3","title":"Radiation Oncology","twitterHandle":"@OncoBioMed","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8c2fb4bf-e2af-401c-9787-74680080f3d2","owner":[],"postedDate":"May 27th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-09T16:02:21+00:00","versionOfRecord":{"articleIdentity":"rs-6632250","link":"https://doi.org/10.1186/s13014-026-02811-2","journal":{"identity":"radiation-oncology","isVorOnly":false,"title":"Radiation Oncology"},"publishedOn":"2026-03-02 15:57:12","publishedOnDateReadable":"March 2nd, 2026"},"versionCreatedAt":"2025-05-27 06:32:03","video":"","vorDoi":"10.1186/s13014-026-02811-2","vorDoiUrl":"https://doi.org/10.1186/s13014-026-02811-2","workflowStages":[]},"version":"v1","identity":"rs-6632250","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6632250","identity":"rs-6632250","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}