Deep learning for fast screening and localization of spinal dural arteriovenous fistulas to enhance clinical workflow

Fei Zheng; Xuyang Cao; Jianmin Xu; Nana Wang; Futao Zhang; Lingling Zhang; Kewei Liang; Li Yang; Qi Guo; Yali Wang; Ping Yin; Xuedan Feng; Xuzhu Chen; Nan Hong

doi:10.1038/s41746-026-02474-z

. 2026 Feb 27;9:296. doi: 10.1038/s41746-026-02474-z

Deep learning for fast screening and localization of spinal dural arteriovenous fistulas to enhance clinical workflow

Fei Zheng ^1,^#, Xuyang Cao ^1,^2,^#, Jianmin Xu ², Nana Wang ³, Futao Zhang ², Lingling Zhang ⁴, Kewei Liang ⁵, Li Yang ⁶, Qi Guo ², Yali Wang ², Ping Yin ^1,^✉, Xuedan Feng ^7,^✉, Xuzhu Chen ^4,^8,^✉, Nan Hong ^1,^✉

PMCID: PMC13065795 PMID: 41760890

Abstract

Manual processing of spinal CTA images for detecting dural arteriovenous fistulas (SDAVF) is laborious and operator-dependent. We developed an automated AI system (SDAVFdoc) that integrates a 3D convolutional neural network with anatomical prior knowledge to both identify SDAVF and localize the fistula site in a multicenter study of 718 patients. The system sequentially segments spinal structures, localizes the fistula region using anatomical priors, and finally classifies SDAVF likelihood via DenseNet within the foramina. The draining vein cluster segmentation model, using a threshold of 42.5, achieved high accuracy in distinguishing SDAVF cases, with F1-scores ranging from 0.932 to 0.960 across test sets. The DenseNet-based fistula detection model showed a high AUC of 0.928-0.954 across test sets. Compared to technicians, the system reduced processing time from 40.75 ± 11.57 min to 1.05 ± 0.29 min (P < 0.001) and clicks from 761.80 ± 202.05 to 9.68 ± 2.12 (P < 0.001), greatly streamlining clinical workflows. This AI-driven approach enables fast, accurate screening and localization of SDAVF.

Subject terms: Anatomy, Computational biology and bioinformatics, Engineering, Health care, Mathematics and computing, Medical research

Introduction

Spinal dural arteriovenous fistulas (SDAVFs) represent the majority of spinal vascular malformations, accounting for approximately 70% of such cases¹. Although SDAVFs represent the most prevalent type of spinal vascular malformation, they are frequently underdiagnosed due to a combination of their relative rarity, nonspecific symptoms, diagnostic imaging pitfalls, and limitations in clinical thinking, making their misdiagnosis a systemic issue^2–4. Delayed treatment of SDAVFs may result in irreversible damage to spinal nerve function⁵. Misdiagnosis of this condition can lead to incorrect and potentially harmful treatments¹. Both missed and wrong diagnoses place significant financial and psychological burdens on the patient. The accurate localization of the fistula is crucial for both diagnosing and treating these malformations⁶. Recent advancements in non-invasive vascular imaging techniques have significantly enhanced the diagnosis of SDAVFs⁷. Modalities such as dynamic magnetic resonance angiography has become standard methods for diagnosing SDAVFs^8–10. However, due to the challenges in accurately identifying the fistula site, digital subtraction angiography (DSA) remains the gold standard¹¹.

DSA is considered the gold standard for diagnosing and assessing the detailed angioarchitecture of a SDAVF^1,12. Nevertheless, spinal angiography, when compared to cranial neuroendovascular procedures, is associated with considerably higher radiation exposure¹¹. This is not only due to differences in vascular anatomy but also because the imaging process involves larger body areas with greater volume and thickness¹³. For instance, Luetmer et al. reported an average fluoroscopy time of 38 minutes and a mean contrast agent volume of 219 mL in their series of preoperative spinal angiograms of 22 patients with SDAVFs¹⁴. Moreover, spinal vascular malformations are easily missed during DSA, often necessitating repeat angiography. In conclusion, the numerous injections required in spinal angiography can lead to prolonged fluoroscopy times, greater radiation exposure, and the use of additional contrast agents¹⁵. This highlights the need for non-invasive diagnostic alternatives that reduce radiation risks and supplement traditional diagnostic methods.

Computed tomography angiography (CTA) has emerged as a widely used, minimally invasive, and cost-effective imaging modality for diagnosing vascular disorders^16,17. In the context of SDAVFs, CTA is usually performed prior to DSA to guide the diagnostic process. Typically, CTA is useful in locating the fistula site, thus limiting the need for more invasive angiographic procedures. However, the image reconstruction required for CTA, which depends on the expertise of the CT scan technician, is an operator-dependent, time-consuming, and error-prone task¹⁸. The accuracy of the reconstruction is directly influenced by the operator’s experience, and such limitations can be particularly problematic in cases with large patient volumes or during emergency situations^19,20. Delays in the reconstruction process can lead to extended patient triage times, treatment delays, and increased burden on post-processing staff. Given the increasing demand for CTA examinations, this challenge is becoming more pronounced, highlighting the urgent need for an automated and efficient method to identify and locate SDAVFs while minimizing errors and alleviating the workload of operators.

Artificial intelligence (AI), encompassing various computational models and algorithms designed to address complex real-world problems, has made significant strides in the field of radiomics and medical imaging^21,22. Deep neural networks, particularly convolutional neural networks (CNNs), have emerged as effective tools for tackling the technical challenges of medical imaging^23–25. U-Net, a specialized CNN, has shown outstanding performance in biomedical image segmentation^26,27. In brief, AI’s advancements in automating medical image post-processing tasks have resulted in substantial reductions in time and effort required for such processes^28,29. Despite these advances, however, no intelligent diagnostic or recognition system for whole spinal cord CTA has been developed to date. We hypothesize that two main factors contribute to this: the rarity of cases and the complexity of the task. Firstly, SDAVFs are rare, and most publications on the subject are case reports or literature reviews^30–32. China has nearly 3275 top-tier hospitals, but few can perform whole-spinal CTA³³. Secondly, the identification of SDAVFs remains particularly challenging. Unlike vascular structures in the head, neck, and other regions, SDAVFs are not only small in size but also technically demanding to process³⁴. This difficulty is clearly illustrated by our dataset, which shows that the draining vein cluster occupies merely 0.00185% of the total imaging area (Supplementary Table 1). To address these challenges, we developed a multi-stage workflow specifically tailored for SDAVF analysis. The development of such a system could significantly streamline the diagnosis and management of SDAVFs.

In this study, we developed the SDAVFdoc system, which can automatically screen for patients with SDAVF and further automatically locate the fistula site in the identified patients. We conducted a thorough quantitative evaluation of SDAVFdoc, assessing the model’s algorithmic performance, the quality of image reconstruction in meeting diagnostic standards, and the efficiency of its clinical application. We believe that AI technology, when integrated into the radiology workflow, can enhance operational efficiency, reduce healthcare costs, and ultimately improve patient outcomes.

Results

Clinical characteristics

The average age of all participants was 55 ± 13 years, with a range from 15 to 82 years. Specifically, the average age of patients with positive DAVF was 61 ± 11 years, while the average age of patients with negative DAVF was 50 ± 13 years. Among the total cohort, 543 patients (75.6%) were male, and 175 patients (24.4%) were female. Within the group of patients with positive DAVF, 287 (79.5%) were male, whereas among those with negative DAVF, 256 (71.7%) were male. In terms of the distribution of fistula locations, among the patients with positive DAVF, 173 patients (47.9%) had fistulas located in the thoracic segment, while 188 patients (52.1%) had fistulas in the lumbar segment. The clinical characteristics of the patients included in the internal and external datasets are provided in Table 1.

Table 1.

Baseline characteristics of the patients

	Internal set	Internal set (+)	Internal set (−)	P-value	External set	External set (+)	External set (−)	P-value
Age	54.96 ± 12.77	60.25 ± 10.87	50.38 ± 12.55	0.001	54.75 ± 13.10	62.45 ± 10.40	47.91 ± 11.39	0.001
Sex				0.112				0.887
Female	147 (23.9%)	58 (19.3%)	89 (28.3%)		28 (27.5%)	16 (26.7%)	12 (28.6%)
Male	469 (76.1%)	243 (80.7%)	226 (71.7%)		74 (72.5%)	44 (73.3%)	30 (71.4%)

Open in a new tab

Note: (+) denotes the cohort of patients with confirmed SDAVF; (–) indicates the cohort of patients without SDAVF.

SDAVFdoc model performance

ResUnet2, which segments draining vein clusters, achieved Dice coefficients of 0.802 on the validation set, 0.795 on the retrospective internal test set, 0.784 on the prospective internal test set, and 0.785 on the external test set. ResUnet1, dedicated to spinal cord segmentation, obtained Dice scores of 0.946, 0.940, 0.937, and 0.939 across the same four sets, respectively. ResUnet3, which identifies vertebral body centroids, yielded Dice values of 0.901, 0.897, 0.886, and 0.888. ResUnet4, designed for intervertebral foramen recognition, produced Dice coefficients of 0.921, 0.916, 0.917, and 0.920 on these datasets. Complete results are provided in Table 2.

Table 2.

Quantitative evaluation of the segmentation model’s performance

Sets	Dice coefficient_ResUnet1	Dice coefficient_ResUnet2	Dice coefficient_ResUnet3	Dice coefficient_ResUnet4
Validation set	0.946	0.802	0.901	0.921
Retrospective internal test set	0.940	0.795	0.897	0.916
Prospective internal test set	0.937	0.784	0.886	0.917
External test set	0.939	0.785	0.888	0.920

Open in a new tab

On the validation set, we assessed how ResUnet2’s performance in segmenting draining vein clusters varied with pixel value thresholds (from 20 to 80 in increments of 10), as shown in Fig. 1 and Supplementary Table 2. The highest F1 score was observed at a threshold of 40. Further analysis indicated that the model’s performance reached its optimal F1 value of 0.949 within the segmentation pixel value threshold range of 41 to 44, with identical confusion matrices across these thresholds. Consequently, an average threshold value of 42.5 was selected. At this threshold, ResUnet2’s performance was evaluated across various test sets (Table 3). On the validation set, the model demonstrated strong performance with an accuracy of 0.956, precision of 0.942, recall of 0.970, and an F1 score of 0.956. Similarly, on the retrospective internal test set, it achieved an accuracy of 0.955, a precision of 0.927, a recall of 0.984, and an F1 score of 0.955. Performance remained consistently high on the prospective internal test set, with values of 0.964 in accuracy, 1.000 in precision, 0.923 in recall, and 0.960 in F1 score. For the external test set, the model attained an accuracy of 0.922, precision of 0.948, recall of 0.917, and an F1 score of 0.932. Confusion matrices illustrating the model’s performance across various datasets (validation set, retrospective internal test set, prospective internal test set, and external test set) are shown in Supplementary Fig. 1.

Fig. 1 — a Model performance metrics (accuracy, precision, recall, F1-score) across different voxel value thresholds (20 to 80 in increments of 10). The line plot shows the trend of the F1-score. b Confusion matrices for the model at the corresponding thresholds shown in (a).

Table 3.

Performance evaluation of ResUnet2 across validation and test sets at optimal threshold (threshold value = 42.5)

Sets	Confusion matrix	Accuracy	Precision	Recall	F1-score
Validation set	[65, 4],	0.956	0.942	0.970	0.956
Validation set	[2, 65]	0.956	0.942	0.970	0.956
Retrospective internal test set	[64, 5],	0.955	0.927	0.984	0.955
Retrospective internal test set	[1, 63]	0.955	0.927	0.984	0.955
Prospective internal test set	[15, 0],	0.964	1.000	0.923	0.960
Prospective internal test set	[1, 12]	0.964	1.000	0.923	0.960
External test set	[39, 3],	0.922	0.948	0.917	0.932
External test set	[5, 55]	0.922	0.948	0.917	0.932

Open in a new tab

The detection rates of DenseNet for fistula locations were 0.910 (61/67) on the validation set, 0.922 (59/64) on the retrospective internal test set, 0.846 (11/13) on the prospective internal test set, and 0.883 (53/60) on the external test set, with corresponding specificities of 0.971, 0.971, 1.000, and 0.952 (Table 4). The DenseNet-based model for fistula location detection demonstrated high discriminative performance, with AUC (95% CI) values as follows: 0.954 (0.910–0.989) for the validation set, 0.942 (0.887–0.985) for the retrospective internal test set, 0.928 (0.796–1.000) for the prospective internal test set, and 0.929 (0.862–0.982) for the external test set. Figure 2a illustrates the corresponding ROC curves and confusion matrix heatmaps. Figure 2b demonstrates a representative case where conventional CTA reconstruction failed to clearly visualize the SDAVF, while our SDAVFdoc system successfully localized the fistula at the left 9th intervertebral foramen—a finding subsequently confirmed by DSA as the diagnostic gold standard. This case exemplifies the superior diagnostic capability of our AI-based model in detecting clinically significant vascular abnormalities that may be overlooked in standard CTA imaging protocols.

Table 4.

Performance evaluation of DenseNet across validation and test sets

Sets	Accuracy	Precision	Recall	Specificity
Validation set	0.941	0.968	0.910	0.971
Retrospective internal test set	0.947	0.967	0.922	0.971
Prospective internal test set	0.929	1.000	0.846	1.000
External test set	0.912	0.964	0.883	0.952

Open in a new tab

Fig. 2 — a Discriminative performance of the DenseNet-based fistula detection model, presented as receiver operating characteristic (ROC) curves and confusion matrices. b Comparative imaging from a representative patient with a spinal dural arteriovenous fistula (SDAVF). The computed tomography angiography (CTA) reconstruction image is shown alongside the corresponding digital subtraction angiography (DSA) image, which serves as the ground truth. While the CTA image fails to clearly reveal the fistula, the SDAVFdoc system accurately identifies the outlet at the location of the 9th left intervertebral foramen, as confirmed by DSA.

Clinical evaluation

SDAVFdoc automatically generates draining vein cluster segmentation images from whole spinal cord CTA scans. To further assess its utility, we evaluated whether the generated image quality meets the diagnostic standards expected by radiologists. For the SDAVFdoc, inter-rater agreement in both the retrospective internal and external test sets was substantial, with Cohen’s kappa values of 0.806 and 0.755, respectively. Among the 64 scans in the retrospective internal test set, the overall image pass rate of SDAVFdoc outputs was 92.2% (59/64) for one evaluator and 95.3% (61/64) for the other. In the external test set comprising 60 scans, the pass rate was 90.0% (54/60) according to one evaluator and 93.3% (56/60) for the second evaluator. Notably, images rated as excellent quality (score of 3) accounted for more than 78.1% (50/64) of the retrospective internal test set and more than 78.3% (47/60) of the external test set. The image quality evaluation showed that 3-point ratings corresponded to large, well-defined draining vein clusters, 2-point ratings indicated irregular draining vein clusters with slightly larger volumes, and 1-point ratings represented smaller, isolated draining vein clusters (Fig. 3). These findings reflect a clear distinction between various levels of image quality based on the size and morphology of the draining vein clusters.

Fig. 3 — The upper image shows the distribution of the 1 to 3-point ratings in the overall image quality assessments dataset. The lower image presents representative examples for each of the 1 to 3-point ratings: In the 3-point image, the draining vein cluster has a regular shape and large volume; in the 2-point image, the draining vein cluster has an irregular shape and slightly larger volume; in the 1-point image, the draining vein cluster is a single vessel with a smaller volume.

The average post-processing time for the two technicians was 39.46 ± 11.64 min and 42.04 ± 12.06 min, respectively. In comparison, the average processing time for SDAVFdoc was significantly reduced, decreasing from 40.75 ± 11.57 min, the combined average post-processing time for the two technicians, to just 1.05 ± 0.29 min (P < 0.001) (Fig. 4a). Additionally, the average number of clicks required by the two technicians was 717.82 ± 163.94 and 805.79 ± 266.37, respectively. The implementation of SDAVFdoc resulted in a dramatic reduction in the number of clicks, from 761.80 ± 202.05, the average number of clicks for the two technicians, to just 9.68 ± 2.12 (P < 0.001) (Fig. 4b).

Fig. 4 — a Comparison of total processing time (in minutes) between two technicians and the SDAVFdoc system for both SDAVF-positive and negative cases. The upper section displays absolute time usage, while the lower left line plot illustrates the time difference between manual and AI processing, clearly demonstrating a significant reduction in average processing time with AI. An embedded difference plot further emphasizes this reduction, and a heatmap provides detailed differential analysis. b Comparison of the number of operational clicks (counts) required by two technicians and the AI system. The upper section shows click counts, the line plot compares manual versus AI performance, and supplementary difference visualization and heatmap detail the variations. Collectively, the results confirm that the AI system significantly reduces both processing time and manual interaction across diagnostic operations.

Discussions

SDAVF is a vascular anomaly characterized by an abnormal connection between a radicular artery and a radicular vein^1,35. The fistula is usually situated near the spinal nerve root within the dura mater³⁶. The abnormal vascular connection creates a direct conduit for pressurized arterial blood to enter the venous system, leading to venous expansion and congestion. The majority of SDAVFs are solitary lesions arising from the lower thoracic and upper lumbar vertebral segmental arteries³⁷. These anatomical understanding forms the foundation for our experimental design. Additionally, SDAVFs predominantly affect male patients, with diagnosis typically occurring in the fifth or sixth decade of life^38,39. These demographic findings are consistent with the results of our own research.

Delayed diagnosis of SDAVFs is common, with these conditions frequently being underdiagnosed or misdiagnosed as other inflammatory, degenerative, or oncologic spinal disorders in approximately 22% of patients^37,40,41. Catheter-based spinal DSA is the gold standard for diagnosing spinal dural arteriovenous fistulas (SDAVFs) due to its high resolution³⁸. However, DSA is invasive, technically demanding, and carries risks such as radiation exposure, vascular injury, and contrast-related complications⁴². Its accuracy heavily depends on operator expertise, and incomplete catheterization can lead to misdiagnosis⁴³. In contrast, CTA offers a less invasive alternative. When CTA successfully localizes a suspected SDAVF, it enables a targeted DSA approach. This focused strategy simplifies the procedure, shortens exam time, and reduces the risk of the procedure. Despite the advantages of CTA, its postprocessing remains time-consuming and requires a high level of expertise. Such difficulties highlight the demand for noninvasive imaging tools that can both accurately identify SDAVFs and precisely determine the fistula location.

The SDAVFdoc system employs a comprehensive workflow: spinal cord segmentation (ResUnet1) and draining vein cluster segmentation (ResUnet2) to identify the presence of SDAVF, followed by vertebral body centroid identification (ResUnet3) and intervertebral foramen recognition (ResUnet4) with a connected-domain algorithm for localization. Ultimately, fistula recognition is determined using the DenseNet model. The hybrid system integrates these models into a fully automated pipeline, processing whole spinal cord CTA images to detect SDAVF and accurately localize their position. Notably, this study addresses a critical gap in the field, as there is currently a lack of intelligent diagnostic systems specifically designed for SDAVF. Our work represents a significant step toward developing an automated solution for this clinically significant yet underserved diagnostic challenge.

Remarkably, we also systematically assessed SDAVFdoc’s clinical utility by evaluating its diagnostic image quality and workflow efficiency enhancement capabilities. For technicians, locating very small blood vessel regions is often time-consuming and heavily reliant on experience. Our method significantly reduces the time required to process each image. Additionally, the tool can accurately pinpoint the location of fistulas, providing complete and comprehensive image information with ease. Therefore, SDAVFdoc has the potential to assist technicians in clinical practice by quickly locating blood vessels (fistulas) in CTA images through an automated process. This innovative workflow for whole-spine CTA patient screening and fistula identification offers numerous advantages, including improved interpretative consistency, time and labor savings, and high sensitivity and specificity. If fully integrated into clinical settings, automated processing of segmented images could simplify the reconstruction process and seamlessly align with the clinical workflow.

Initial attempts to segment draining vein clusters directly from original whole spinal cord CTA images proved unsuccessful, resulting in a Dice coefficient of 0. This failure is likely attributable to the extremely small volumetric representation of the vessels, which led the model to overwhelmingly favor background prediction. To address this, we introduced a preprocessing step in which spinal cord segmentation is performed first, substantially reducing the anatomical search space and improving segmentation feasibility. Quantitative support for this approach, including the area ratios of draining vein clusters and spinal cord relative to the total image, is provided in Supplementary Table 1. Moreover, during the image quality assessment, some images were of poor quality. The main reason for this was the relatively small volume proportion of the draining vein cluster. This further highlights the significance of the proportion of the draining vein cluster within the image area in our experiment.

Although the results of this study are promising, there are several limitations. Firstly, the datasets exclusively comprised cases of SDAVF, omitting other spinal vascular malformations such as extradural or perimedullary arteriovenous fistulas, which restricts the model’s current applicability and necessitates future validation on a more diverse spectrum of pathologies. Secondly, the model’s development was specifically optimized for thoracolumbar (T1-L5) CTA imaging protocols, meaning its performance for fistulas outside this region remains unproven and requires dedicated assessment. Thirdly, our cohort did not include rare but clinically challenging cases of SDAVF with multiple fistulous points, leaving the model’s efficacy for such complex variants unknown. Finally, while the AI system demonstrates high diagnostic accuracy, its decision-making process as a deep learning model lacks inherent interpretability, which could pose a challenge for clinical trust and adoption without further development of explainability features. We acknowledge that these limitations do not invalidate our core findings but clearly delineate the scope for future work to enhance the model’s robustness, versatility, and translational potential.

In conclusion, we have successfully developed a CNN model capable of detecting and localizing SDAVF through a fully automated process using whole-spine CTA images. Compared to existing technologies, it provides a time-efficient and objective approach to optimize the segmentation and identification of whole-spine CTA images, thereby reducing costs and improving efficiency. The model demonstrates high reproducibility and generalizability, making it a potentially valuable tool for the non-invasive characterization of SDAVF. The efficiency and accuracy demonstrated by SDAVFdoc in this study may support its future consideration as a supportive tool in clinical workflows.

Methods

This study was conducted in accordance with the Declaration of Helsinki. The study protocol was reviewed and approved by the Institutional Review Boards/Ethics Committees of all participating centers: the Ethics Committee of Beijing Fengtai You’ anmen Hospital (Approval No.: KY-202405-01 for the retrospective component and KY-202501-01 for the prospective component), the Ethics Committee of Beijing Haidian Hospital (Haidian Campus of Peking University Third Hospital) (Approval No.: YYLL202201235V), the Ethics Committee of Beijing Tiantan Hospital, Capital Medical University (Approval No.: KY-2021-012-02). Informed consent was waived for the retrospective phase, whereas written informed consent was obtained from all participants in the prospective phase.

Study design

As shown in Fig. 5, our automated hybrid system, SDAVFdoc, employs a three-stage architecture incorporating five CNNs.

The first stage determines whether the patient has an SDAVF. ResUnet1 is used for spinal cord segmentation, followed by expanding the spinal cord region outward by six voxels to create an expansion zone. Only the expanded region is retained, and the region is set to a soft tissue window. Using ResUnet2, the draining vein cluster, specifically the perimedullary veins, is automatically segmented. The resulting mask is then resized to the original image size. The volume of the perimedullary veins is calculated, and if the volume is below a threshold, the patient is considered not to have an SDAVF. Conversely, if the volume exceeds the threshold, the patient is identified as having an SDAVF.

In the second stage, to narrow down the potential anatomical region of the fistula, a three-class segmentation approach using ResUnet3 is applied to the CT images, where Class 0 represents the background, Class 1 represents spherical regions centered on the vertebral body centroids from T1 to L5, and Class 2 represents the sacral region. A connected component labeling algorithm is used to distinguish the spherical regions centered on the vertebral body centroids from T1 to L5. The z-axis coordinates of each spherical region and the sacral vertebral centroid are obtained. Combining this anatomical prior knowledge, the sacral region is located below the L5 vertebra, and the vertebrae above L5 are L4, L3, L2, L1, extending up to T1. By matching the z-axis coordinates of the spherical regions with the vertebral names, the ZT1-ZL5 values are determined. The highest z-axis coordinate of the draining veins (Zhigh) and the lowest coordinate (Zlow) are identified. Among the ZT1-ZL5 17 values, the smallest value greater than Zhigh and the largest value less than Zlow are used to crop the CT region between ZTn and ZTn + x, which represents the potential anatomical region of the fistula.

In the third stage, for fistula localization, the potential anatomical region of the fistula is divided into smaller blocks based on the coordinate values, ranging from ZTn to ZTn + 1, ZTn + 1 to ZTn + 2, and so on. ResUnet4 is applied to each small block to segment spherical regions centered on the bilateral intervertebral foramina and compute the centroids of these regions. The left-side (Lx, Ly, Lz) and right-side (Rx, Ry, Rz) coordinates of the centroids are obtained. A 32 × 32 × 32 CT region is cropped around the intervertebral foramen. Finally, DenseNet is used to evaluate the probability of the fistula being present in each 32 × 32 × 32 CT region. To further reduce the false positive rate, patients with output probabilities below 0.2 for any potential anatomical region of the fistula are predicted as negative. For patients with probabilities greater than 0.2, in order to enhance the detection rate, the CT region with the highest probability is identified as the region containing the fistula. The ZTn + x − ZTn + x + 1 range, along with the left or right coordinates, is then used to determine the precise location of the fistula within the intervertebral foramen.

Patient data and CTA images acquisition

All whole spinal cord CTA scans were performed as part of routine clinical care. The scanning range extended from the upper edge of the eye socket to the end of the coccygeal vertebra, with a total of 1198–1205 slices along the z-axis. The acquired patient CTA images were in DICOM format and were obtained using GE and Philips scanners across multiple centers. The specific scanning parameters for Center 1 are detailed in Supplementary Table 3. The imaging protocol for Centers 2 and 3 (using a Philips scanner) is described as follows: CTA was performed using a Philips 128-slice CT scanner. The scanning parameters were set as follows: tube voltage 120 kV, automatic tube current modulation with a tube current–time product of 200–300 mAs, slice thickness 0.5 mm, pitch 0.899, and gantry rotation time 0.5 s per rotation. The scanning range extended from the skull base to the sacrococcygeal region, covering the entire spinal cord. A bolus of contrast agent (ioversol, 320 mgI/mL) was intravenously administered via the antecubital vein using a power injector at a flow rate of 4 mL/s. The dose was adjusted according to body weight at 1.5 mL/kg, with a maximum total volume not exceeding 100 mL. Bolus-tracking technique was employed with the descending aorta as the region of interest; scanning was automatically triggered once the attenuation reached a threshold of 100 Hounsfield units. All images were acquired with a 512×512 matrix and transferred to a workstation for post-processing. Reconstruction techniques included volume rendering (VR), multiplanar reformation (MPR), and maximum intensity projection (MIP). Iopromide, a contrast agent with a concentration of 370 mgI/mL, was used at an injection rate of 4.5 mL/s.

Between July 2016 and January 2025, a total of 718 patients were enrolled across three participating centers, comprising 361 SDAVF-positive and 357 SDAVF-negative cases, confirmed by DSA and/or surgical findings. The primary cohort, used for model development, included 224 positive and 231 negative patients. Model evaluation was performed using retrospectively collected internal and external test sets, while a prospective set was employed for clinical validation (Fig. 6). The distribution of data across training, validation, and test subsets is illustrated in Supplementary Fig. 2. Between July 2016 and January 2025, Center 1 (a specialized referral center for SDAVF) enrolled a total of 616 patients. Among them, 301 were confirmed as SDAVF-positive by DSA and surgical results, while the remaining 315 were SDAVF-negative. Similarly, Center 2 and Center 3 collectively enrolled 102 patients, including 60 SDAVF-positive and 42 SDAVF-negative cases, all confirmed by DSA and/or surgical outcomes. The primary cohort from center 1, used for model training, hyperparameter optimization, and validation, consisted of 268,374 images from 224 SDAVF-positive patients (231 SDAVF-negative patients) between July 2016 and April 2024. A retrospective internal test set (Center 1) and an external test set (Center 2,3), used for model testing and image quality assessment, included 76,665 images from 64 SDAVF-positive patients (69 SDAVF-negative patients) enrolled at Center 1 between May 2024 and October 2024, and 71,887 images from 60 SDAVF-positive patients (42 SDAVF-negative patients) enrolled at Center 2,3. Additionally, the prospective test set group, used for model testing and clinical efficiency evaluation, included 15,377 images from 13 SDAVF-positive patients (15 SDAVF-negative patients) between November 2024 and January 2025.

Fig. 6 — The process is divided into three parts: a training set, a validation set, and three test sets (one being a prospective independent test set). Four segmentation or identification models are constructed using ResUnet, and a fistula identification model is constructed using DenseNet, with performance evaluations conducted for each model. Image quality assessments and clinical performance evaluations are carried out based on three independent test sets, with accuracy analysis used to validate the reliability of our model. “P” denotes SDAVF-positive patients; “N” denotes SDAVF-negative patients.

The inclusion and exclusion criteria were consistent across all three participating centers. Patients with suspected SDAVF underwent CTA. This cohort comprised 616 patients from Center 1 and 102 patients from Centers 2 and 3. The inclusion criteria were as follows: (1) patients who underwent spinal cord CTA examination; (2) patients with complete imaging and clinical data; (3) the presence or absence of SDAVF was definitively confirmed using DSA or surgical findings. The exclusion criteria included: (1) incomplete imaging data; (2) poor image quality; (3) patients with recurrence of SDAVF after surgery or those requiring repeated follow-up or monitoring.

Two blinded, experienced radiologists independently assessed images in the retrospective sets using a 3-point scale to evaluate diagnostic suitability. The prospective set was designed primarily to evaluate model performance under real-world clinical conditions. It also incorporated assessments of workflow efficiency, including time savings and the number of post-processing operations required.

Radiologist’s annotations

Given the substantial dataset, manual segmentation was conducted on scans from 83 patients diagnosed with SDAVF, with the remaining cases processed using automatic segmentation. A randomly selected subset of 83 SDAVF-positive patients from the primary cohort served as the annotated dataset. The annotation process was performed by radiologists with a standardized protocol using ITK-SNAP 3.8.0 (http://www.itksnap.org): for ResUnet1, spinal cord labels were generated through direct manual segmentation on whole-spine CTA images; for ResUnet2, draining vein cluster annotations were created within an anatomically constrained region derived from the spinal cord mask dilated by 6 voxels; while for both ResUnet3 and ResUnet4, reference labels consisted of standardized spherical regions centered on vertebral body centroids and bilateral intervertebral foramina centroids, respectively, as identified by radiological experts. Algorithm-generated annotations were reviewed and corrected by radiologists.

To minimize human errors, we adopted a hierarchical annotation method for all scanned images in the database. Initially, the dataset was pre-annotated using ITK-SNAP, where the spinal cord, draining vein cluster (for SDAVF-positive patients), the center of the vertebral body region, and the bilateral intervertebral foramina were marked with different masks. Subsequently, two technicians with more than two years of post-processing experience independently corrected the pre-annotated images to generate the initial annotation results. Afterward, two radiologists with over five years of experience rechecked the accuracy of the initial annotations and made corrections when necessary to ensure the precision of the annotations. In cases where the opinions of the two radiologists differed, a senior arbitration expert, a radiologist with over ten years of experience, made the final decision. The average time required for pre-annotating each case was approximately 47 ± 17 minutes, with the time for annotating the draining vein cluster being 38 ± 14 minutes.

SDAVFdoc model development

The SDAVFdoc model is an automated hybrid system that employs five CNNs, including four segmentation models based on ResUnet and one classification model based on DenseNet, for segmenting and recognizing spinal structures and the draining vein cluster. It is specifically designed to detect and localize SDAVFs in whole spinal cord CTA images.

The ResUnet-based segmentation models were developed as follows. Initially, we resampled the voxel size of all images to (0.5, 0.5, 0.625). Given the large size of the images, it was not feasible to input them in their entirety into the segmentation network. Therefore, we adopted a patch-based processing approach. The patch sizes were as follows: [160, 160, 320] for the spinal cord, [80, 80, 700] for the draining vein cluster, [144, 144, 400] for the vertebral bodies, and [120, 120, 80] for the intervertebral foramina. Since draining vein clusters are more prominent under soft tissue window settings, the images were adjusted to soft tissue window mode to enhance feature visibility. To improve dataset diversity and the model’s robustness, we applied random flipping to both the images and their corresponding labels. Specifically, a 50% probability of flipping was applied along each of the three spatial dimensions (x, y, and z). This method enabled the model to learn variations in image orientation and generalize better across different spatial transformations.

The DenseNet-based classification model was trained and evaluated as follows. The primary cohort was randomly partitioned into training and validation subsets at a 7:3 ratio. Model evaluation was performed using two independent test cohorts: a retrospective dataset and a prospectively collected test set. To enhance the model’s robustness and generalization ability, we implemented several data augmentation techniques on the training images. These techniques included random rotation, random flipping, and random scaling. For random rotation, the images were rotated within a range of [−π/12, π/12] with a 50% probability, simulating variations in image orientation to help the model adapt to different angles. Random flipping was applied along each of the three spatial dimensions (x, y, and z) with a 50% probability, enabling the model to learn spatial invariance. Additionally, random scaling was performed within a range of 90% to 100% of the original size, with a 50% probability, allowing the model to handle variations in image size and improving its robustness to scale changes.

The ResUnet and DenseNet models together form a hybrid system called SDAVFdoc, which features a fully automated workflow that ensures seamless data transfer between components. In this study, the Adam optimizer (with a learning rate of 1e-4) was used to train the ResUNet/DenseNet model and optimize its parameters. To control the learning rate more effectively, we employed the WarmupCosineSchedule learning rate scheduler. This scheduler initiated three warm-up steps at the beginning of training, gradually increasing the learning rate from 0.1 times the initial value to the set base value of 1e-4. Afterward, it used a cosine annealing strategy to progressively decrease the learning rate until the end of training. This scheduling method facilitated smooth convergence during training and improved the overall performance of the model. The model was trained for 1,000 epochs using ResUnet and for 100 epochs using DenseNet. All computational experiments were conducted using Python 3.8 with PyTorch 1.9.0 framework, deployed on a Linux-based workstation equipped with an NVIDIA RTX A6000 GPU (48 GB GDDR6 memory)

Together, these models form a hybrid system, SDAVFdoc, with a fully automated pipeline that ensures seamless data transfer between components. The input data comprises integrated, continuous three-dimensional images of whole-spine CTA. The model’s output provides a classification of whether the patient has an SDAVF, along with the specific location of the fistula (e.g., at the right intervertebral foramen of L3).

Clinical evaluation

The clinical evaluation of the SDAVFdoc system encompassed both image quality assessments and an analysis of clinical efficiency.

The automated SDAVFdoc is designed to provide an efficient, quality-assured, and unattended image recognition and diagnostic process. With the reconstruction of draining vein clusters directly impacting the subsequent fistula localization, accurate segmentation of these draining vein clusters is critical for proper diagnosis, and the segmented images are evaluated by two experienced radiologists (with more than 8 years of experience) to determine whether they meet clinical diagnostic standards.

For the image quality assessment, a retrospective internal test set and an external test set were utilized to calculate the image pass rate, with subjective evaluation scores assigned as follows: a score of 3 represents excellent image quality, with complete vascular delineation, no artifacts, and easy diagnosability; a score of 2 indicates good image quality, where the vascular delineation is relatively complete but some vessels remain unclear, with minimal artifacts, though still sufficient for diagnosis; and a score of 1 denotes poor image quality, where most vessels are unclear, severe artifacts are present, and diagnosis is difficult. Images achieving a minimum score of 2 were considered to fulfill diagnostic quality standards and were consequently classified as passing. In the event of disagreement between the two evaluators, a consensus was achieved through joint reevaluation.

The clinical efficiency of SDAVFdoc was evaluated between November 2024 and January 2025, focusing on two main aspects: time consumption (work duration) and number of clicks (labor intensity). Specifically, during the technician’s operation, another author (CXY) recorded the time and tracked the number of clicks in real-time by manually drawing tally marks, which were summed up afterwards. A prospective cohort of 28 patients (15 with negative SDAVF and 13 with positive SDAVF) was included in the study, and the time spent throughout the entire post-processing workflow, including the reconstruction of draining vein clusters and identification of potential fistula sites by two technicians, as well as the number of clicks involved, was recorded. Similarly, the time and number of clicks required for SDAVFdoc to analyze the same 28 patients were documented. The average results from the two technicians were compared with those of SDAVFdoc, allowing for an assessment of the model’s impact on clinical efficiency, particularly in terms of reducing time consumption and labor intensity, thereby improving the overall diagnostic workflow.

Statistical analysis

Comparative analyses of clinical characteristics among study groups were performed using standardized statistical approaches. Categorical variables were assessed using Pearson’s chi-square test. Normally distributed continuous variables were compared using independent two-sample t-tests or paired t-tests, as appropriate. All statistical tests were conducted as two-sided, with statistical significance defined by a p-value of less than 0.05. The analysis was performed using SPSS 26 (SPSS Inc., Armonk, NY). For model performance evaluation, we implemented a comprehensive assessment framework incorporating multiple metrics: (1) standard classification metrics, including accuracy, precision, recall, specificity, F1-score, and the area under the ROC curve (AUC), to evaluate discriminatory capability; (2) confusion matrices for detailed accuracy quantification; and (3) Dice coefficient⁴⁴ as a specialized metric for spatial overlap accuracy in segmentation tasks. This multi-dimensional evaluation strategy ensured robust assessment of both classification and segmentation performance.

Supplementary information

Supplementary^{(400.8KB, pdf)}

Acknowledgements

This study was supported by the National Natural Science Foundation of China (NO.82471950); Beijing Natural Science Foundation (Grant No. L242061).

Author contributions

F.Z. designed the study, collected and analyzed the data, and prepared and edited the paper. X.Y.C. participated in the clinical research process and edited the paper. J.M.X. participated in the clinical research process and edited the paper. N.N.W. participated in the clinical research process. F.T.Z. Revised the paper. L.L.Z. participated in the clinical research process. K.W.L. designed the study, collected and analyzed the data. L.Y. participated in the clinical research process. Q.G. participated in the clinical research process. Y.L.W. participated in the clinical research process. P.Y. took part in the processing of data analysis and statistics. X.D.F. took part in the processing of data analysis and statistics. X.Z.C. participated in the clinical guidance and article preparation processes. N.H. designed the study, ensured the integrity of the whole study and revised the paper. All authors read and approved the final manuscript.

Data availability

Due to data protection restrictions, the full dataset is not publicly available. However, reasonable inquiries regarding the data can be directed to the corresponding author.

Code availability

The code developed in this study is available in our GitHub repository: https://github.com/Kewei-Liang/SDAVFdoc.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Fei Zheng, Xuyang Cao.

Contributor Information

Ping Yin, Email: yinping915@163.com.

Xuedan Feng, Email: xuedan_feng@163.com.

Xuzhu Chen, Email: radiology888@aliyun.com.

Nan Hong, Email: hongnan1968@163.com.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-026-02474-z.

References

1.Alkhaibary, A. et al. Spinal dural arteriovenous fistula: a comprehensive review of the history, classification systems, management, and prognosis. Chin. Neurosurg. J.10, 2 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
2.Auricchio, A. M. et al. Digital exoscope versus surgical microscope in spinal dural arteriovenous fistula surgery: a comparative series. Neurosurg. Focus56, E13 (2024). [DOI] [PubMed] [Google Scholar]
3.Tanaka, T. et al. Delayed diagnosis of spinal dural arteriovenous fistula: a case report and scoping review. J. Clin. Med.13, 711 (2024). [DOI] [PMC free article] [PubMed]
4.Ouyang, H. et al. Spinal dural arteriovenous fistula with low back pain as the only symptom at onset: a case report. Am. J. Case Rep.25, e943947 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Wang, P. et al. Dural arteriovenous fistula with spinal dural arteriovenous fistula: a case report and review of the literature. J. Med. Case Rep.17, 467 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.He, N., Lyu, H., Lu, Y., Yan, F. & Jiang, H. Accurate detection of spinal dural arteriovenous fistula with spinal photon-counting computed tomography angiography: a report of two cases. Stroke56, e44–e46 (2025). [DOI] [PubMed] [Google Scholar]
7.Marcus, J. et al. Spinal dural arteriovenous fistulas: a review. Curr. Atheroscler. Rep.15, 335 (2013). [DOI] [PubMed] [Google Scholar]
8.Yamaguchi, S. et al. The position and role of four-dimensional computed tomography angiography in the diagnosis and treatment of spinal arteriovenous fistulas. World Neurosurg.103, 611–619 (2017). [DOI] [PubMed] [Google Scholar]
9.Saraf-Lavi, E. et al. The role of MRA as a preliminary diagnostic tool in diagnosing and localizing spinal dural arteriovenous fistula (SDAVF). AJNRAm. J. Neuroradiol.46, 2414–2419 (2025). [DOI] [PMC free article] [PubMed]
10.Meijer, F. J. A., Raymaekers, V., Pegge, S. A. H. & Boogaarts, J. H. Diagnostic accuracy of 4D-MRA for the detection and localization of spinal dural arteriovenous fistulas: a retrospective 10-year cohort study. Brain Spine5, 104176 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Saindane, A. M., Boddu, S. R., Tong, F. C., Dehkharghani, S. & Dion, J. E. Contrast-enhanced time-resolved MRA for pre-angiographic evaluation of suspected spinal dural arterial venous fistulas. J. Neurointervent. Surg.7, 135–140 (2015). [DOI] [PubMed] [Google Scholar]
12.Khalafallah, A. M. et al. Evaluating the diagnostic accuracy of 3D contrast-enhanced magnetic resonance angiography versus digital subtraction angiography in spinal dural arteriovenous fistulas. Neurosurg. focus56, E10 (2024). [DOI] [PubMed] [Google Scholar]
13.Miller, D. L., Kwon, D. & Bonavia, G. H. Reference levels for patient radiation doses in interventional radiology: proposed initial values for U.S. practice. Radiology253, 753–764 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Luetmer, P. H. et al. Preangiographic evaluation of spinal dural arteriovenous fistulas with elliptic centric contrast-enhanced MR Angiography and effect on radiation dose and volume of iodinated contrast material. Am. J. Neuroradiol.26, 711–718 (2005). [PMC free article] [PubMed] [Google Scholar]
15.Ozpeynirci, Y. et al. Radiation dose and fluoroscopy time of diagnostic angiography in patients with spinal dural arteriovenous fistula. Clin. Neuroradiol.32, 791–797 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Powers, W. J. et al. Guidelines for the early management of patients with acute ischemic stroke: 2019 update to the 2018 guidelines for the early management of acute ischemic stroke: a guideline for healthcare professionals from the American Heart Association/American Stroke Association. Stroke50, e344–e418 (2019). [DOI] [PubMed] [Google Scholar]
17.Li, K. et al. A comparative study of AI-based automated and manual postprocessing of head and neck CT angiography: an independent external validation with multi-vendor and multi-center data. Neuroradiology66, 1765–1780 (2024). [DOI] [PubMed] [Google Scholar]
18.Fu, F. et al. Rapid vessel segmentation and reconstruction of head and neck angiograms using 3D convolutional neural network. Nat. Commun.11, 4829 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Fu, F. et al. Deep learning for head and neck CT angiography: stenosis and plaque classification. Radiology307, e220996 (2023). [DOI] [PubMed] [Google Scholar]
20.Sheth, S. A. et al. Machine learning and acute stroke imaging. J. Neurointervent. Surg.15, 195–199 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Tan, P., Chen, X., Zhang, H., Wei, Q. & Luo, K. Artificial intelligence aids in development of nanomedicines for cancer management. Semin. Cancer Biol.89, 61–75 (2023). [DOI] [PubMed] [Google Scholar]
22.Nam, D., Chapiro, J., Paradis, V., Seraphin, T. P. & Kather, J. N. Artificial intelligence in liver diseases: Improving diagnostics, prognostics and response prediction. JHEP Rep. Innov. Hepatol.4, 100443 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Zhang, Z. et al. Development of an MRI based artificial intelligence model for the identification of underlying atrial fibrillation after ischemic stroke: a multicenter proof-of-concept analysis. EClinicalMedicine81, 103118 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Chao, B. T. et al. Improving prognostic accuracy in lung transplantation using unique features of isolated human lung radiographs. NPJ Digit. Med.7, 272 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Zhang, Y. & Chung, A. C. S. Retinal vessel segmentation by a transformer-U-net hybrid model with dual-path decoder. IEEE J. Biomed. Health Inform.28, 5347–5359 (2024). [DOI] [PubMed] [Google Scholar]
26.Wang, Z., Jia, L. V. & Liang, H. Partial class activation mapping guided graph convolution cascaded U-Net for retinal vessel segmentation. Comput. Biol. Med.178, 108736 (2024). [DOI] [PubMed] [Google Scholar]
27.Chen, J. et al. TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers. Med. image Anal.97, 103280 (2024). [DOI] [PubMed] [Google Scholar]
28.Li, M. et al. Deep learning segmentation and reconstruction for CT of chronic total coronary occlusion. Radiology306, e221393 (2023). [DOI] [PubMed] [Google Scholar]
29.van der Ven, J. P. G. et al. Multivendor evaluation of automated MRI postprocessing of biventricular size and function for children with and without congenital heart defects. J. Magn. Reson. Imaging58, 794–804 (2023). [DOI] [PubMed] [Google Scholar]
30.Sturiale, C. L. et al. Spinal dural arteriovenous fistulas: a retrospective analysis of prognostic factors and long-term clinical outcomes in the light of the recent diagnostic and technical refinements. Acta Neurochir. Suppl.135, 223–230 (2023). [DOI] [PubMed] [Google Scholar]
31.Morofuji, Y. et al. Non-sinus type dural arteriovenous fistula: others. J. Neuroendovasc. Ther.19, 2023–0023 (2025). [DOI] [PMC free article] [PubMed]
32.Hiramatsu, M. et al. Non-sinus-type dural arteriovenous fistula at the foramen magnum: a review of the literature. J. Neuroendovasc. Ther.19, 2023–0019 (2025). [DOI] [PMC free article] [PubMed]
33.Chen, L. et al. Status of continuous glucose monitoring use and management in tertiary hospitals of China: a cross-sectional study. BMJ Open13, e066801 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Samaniego, E. A. et al. Most promising approaches to improve brain AVM management: ARISE I consensus recommendations. Stroke55, 1449–1463 (2024). [DOI] [PubMed] [Google Scholar]
35.McKeon, A. et al. Pearls & oy-sters: clues for spinal dural arteriovenous fistulae. Neurology76, e10–e12 (2011). [DOI] [PubMed] [Google Scholar]
36.Moosavi, A., Kalapos, P., Church, E. W., Cockroft, K. M. & Thamburaj, K. Localization of spinal dural arteriovenous fistulas from the spatial relationships of perimedullary vessels on standard MRI. J. Neurosurg. Spine40, 389–394 (2024). [DOI] [PubMed] [Google Scholar]
37.Essibayi, M. A., Srinivasan, V. M., Catapano, J. S., Graffeo, C. S. & Lawton, M. T. Spinal dorsal intradural arteriovenous fistulas: natural history, imaging, and management. Neurology101, 524–535 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Kiyosue, H. et al. Angiographic and clinical characteristics of thoracolumbar spinal epidural and dural arteriovenous fistulas. Stroke48, 3215–3222 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Merland, J. J., Riche, M. C. & Chiras, J. Intraspinal extramedullary arteriovenous fistulae draining into the medullary veins. J. Neuroradiol.7, 271–320 (1980). [PubMed] [Google Scholar]
40.Jellema, K., Tijssen, C. C. & van Gijn, J. Spinal dural arteriovenous fistulas: a congestive myelopathy that initially mimics a peripheral nerve disorder. Brain129, 3150–3164 (2006). [DOI] [PubMed] [Google Scholar]
41.Zalewski, N. L. et al. Unique gadolinium enhancement pattern in spinal dural arteriovenous fistulas. JAMA Neurol.75, 1542–1545 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Forbes, G. et al. Complications of spinal cord arteriography: prospective assessment of risk for diagnostic procedures. Radiology169, 479–484 (1988). [DOI] [PubMed] [Google Scholar]
43.Barreras, P. et al. Analysis of 30 spinal angiograms falsely reported as normal in 18 patients with subsequently documented spinal vascular malformations. Am. J. Neuroradiol.38, 1814–1819 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Dice, L. R. Measures of the amount of ecologic association between species. Ecology26, 297–302 (1945). [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary^{(400.8KB, pdf)}

Data Availability Statement

Due to data protection restrictions, the full dataset is not publicly available. However, reasonable inquiries regarding the data can be directed to the corresponding author.

The code developed in this study is available in our GitHub repository: https://github.com/Kewei-Liang/SDAVFdoc.

[CR1] 1.Alkhaibary, A. et al. Spinal dural arteriovenous fistula: a comprehensive review of the history, classification systems, management, and prognosis. Chin. Neurosurg. J.10, 2 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR2] 2.Auricchio, A. M. et al. Digital exoscope versus surgical microscope in spinal dural arteriovenous fistula surgery: a comparative series. Neurosurg. Focus56, E13 (2024). [DOI] [PubMed] [Google Scholar]

[CR3] 3.Tanaka, T. et al. Delayed diagnosis of spinal dural arteriovenous fistula: a case report and scoping review. J. Clin. Med.13, 711 (2024). [DOI] [PMC free article] [PubMed]

[CR4] 4.Ouyang, H. et al. Spinal dural arteriovenous fistula with low back pain as the only symptom at onset: a case report. Am. J. Case Rep.25, e943947 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR5] 5.Wang, P. et al. Dural arteriovenous fistula with spinal dural arteriovenous fistula: a case report and review of the literature. J. Med. Case Rep.17, 467 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR6] 6.He, N., Lyu, H., Lu, Y., Yan, F. & Jiang, H. Accurate detection of spinal dural arteriovenous fistula with spinal photon-counting computed tomography angiography: a report of two cases. Stroke56, e44–e46 (2025). [DOI] [PubMed] [Google Scholar]

[CR7] 7.Marcus, J. et al. Spinal dural arteriovenous fistulas: a review. Curr. Atheroscler. Rep.15, 335 (2013). [DOI] [PubMed] [Google Scholar]

[CR8] 8.Yamaguchi, S. et al. The position and role of four-dimensional computed tomography angiography in the diagnosis and treatment of spinal arteriovenous fistulas. World Neurosurg.103, 611–619 (2017). [DOI] [PubMed] [Google Scholar]

[CR9] 9.Saraf-Lavi, E. et al. The role of MRA as a preliminary diagnostic tool in diagnosing and localizing spinal dural arteriovenous fistula (SDAVF). AJNRAm. J. Neuroradiol.46, 2414–2419 (2025). [DOI] [PMC free article] [PubMed]

[CR10] 10.Meijer, F. J. A., Raymaekers, V., Pegge, S. A. H. & Boogaarts, J. H. Diagnostic accuracy of 4D-MRA for the detection and localization of spinal dural arteriovenous fistulas: a retrospective 10-year cohort study. Brain Spine5, 104176 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR11] 11.Saindane, A. M., Boddu, S. R., Tong, F. C., Dehkharghani, S. & Dion, J. E. Contrast-enhanced time-resolved MRA for pre-angiographic evaluation of suspected spinal dural arterial venous fistulas. J. Neurointervent. Surg.7, 135–140 (2015). [DOI] [PubMed] [Google Scholar]

[CR12] 12.Khalafallah, A. M. et al. Evaluating the diagnostic accuracy of 3D contrast-enhanced magnetic resonance angiography versus digital subtraction angiography in spinal dural arteriovenous fistulas. Neurosurg. focus56, E10 (2024). [DOI] [PubMed] [Google Scholar]

[CR13] 13.Miller, D. L., Kwon, D. & Bonavia, G. H. Reference levels for patient radiation doses in interventional radiology: proposed initial values for U.S. practice. Radiology253, 753–764 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Luetmer, P. H. et al. Preangiographic evaluation of spinal dural arteriovenous fistulas with elliptic centric contrast-enhanced MR Angiography and effect on radiation dose and volume of iodinated contrast material. Am. J. Neuroradiol.26, 711–718 (2005). [PMC free article] [PubMed] [Google Scholar]

[CR15] 15.Ozpeynirci, Y. et al. Radiation dose and fluoroscopy time of diagnostic angiography in patients with spinal dural arteriovenous fistula. Clin. Neuroradiol.32, 791–797 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR16] 16.Powers, W. J. et al. Guidelines for the early management of patients with acute ischemic stroke: 2019 update to the 2018 guidelines for the early management of acute ischemic stroke: a guideline for healthcare professionals from the American Heart Association/American Stroke Association. Stroke50, e344–e418 (2019). [DOI] [PubMed] [Google Scholar]

[CR17] 17.Li, K. et al. A comparative study of AI-based automated and manual postprocessing of head and neck CT angiography: an independent external validation with multi-vendor and multi-center data. Neuroradiology66, 1765–1780 (2024). [DOI] [PubMed] [Google Scholar]

[CR18] 18.Fu, F. et al. Rapid vessel segmentation and reconstruction of head and neck angiograms using 3D convolutional neural network. Nat. Commun.11, 4829 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR19] 19.Fu, F. et al. Deep learning for head and neck CT angiography: stenosis and plaque classification. Radiology307, e220996 (2023). [DOI] [PubMed] [Google Scholar]

[CR20] 20.Sheth, S. A. et al. Machine learning and acute stroke imaging. J. Neurointervent. Surg.15, 195–199 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR21] 21.Tan, P., Chen, X., Zhang, H., Wei, Q. & Luo, K. Artificial intelligence aids in development of nanomedicines for cancer management. Semin. Cancer Biol.89, 61–75 (2023). [DOI] [PubMed] [Google Scholar]

[CR22] 22.Nam, D., Chapiro, J., Paradis, V., Seraphin, T. P. & Kather, J. N. Artificial intelligence in liver diseases: Improving diagnostics, prognostics and response prediction. JHEP Rep. Innov. Hepatol.4, 100443 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR23] 23.Zhang, Z. et al. Development of an MRI based artificial intelligence model for the identification of underlying atrial fibrillation after ischemic stroke: a multicenter proof-of-concept analysis. EClinicalMedicine81, 103118 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR24] 24.Chao, B. T. et al. Improving prognostic accuracy in lung transplantation using unique features of isolated human lung radiographs. NPJ Digit. Med.7, 272 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR25] 25.Zhang, Y. & Chung, A. C. S. Retinal vessel segmentation by a transformer-U-net hybrid model with dual-path decoder. IEEE J. Biomed. Health Inform.28, 5347–5359 (2024). [DOI] [PubMed] [Google Scholar]

[CR26] 26.Wang, Z., Jia, L. V. & Liang, H. Partial class activation mapping guided graph convolution cascaded U-Net for retinal vessel segmentation. Comput. Biol. Med.178, 108736 (2024). [DOI] [PubMed] [Google Scholar]

[CR27] 27.Chen, J. et al. TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers. Med. image Anal.97, 103280 (2024). [DOI] [PubMed] [Google Scholar]

[CR28] 28.Li, M. et al. Deep learning segmentation and reconstruction for CT of chronic total coronary occlusion. Radiology306, e221393 (2023). [DOI] [PubMed] [Google Scholar]

[CR29] 29.van der Ven, J. P. G. et al. Multivendor evaluation of automated MRI postprocessing of biventricular size and function for children with and without congenital heart defects. J. Magn. Reson. Imaging58, 794–804 (2023). [DOI] [PubMed] [Google Scholar]

[CR30] 30.Sturiale, C. L. et al. Spinal dural arteriovenous fistulas: a retrospective analysis of prognostic factors and long-term clinical outcomes in the light of the recent diagnostic and technical refinements. Acta Neurochir. Suppl.135, 223–230 (2023). [DOI] [PubMed] [Google Scholar]

[CR31] 31.Morofuji, Y. et al. Non-sinus type dural arteriovenous fistula: others. J. Neuroendovasc. Ther.19, 2023–0023 (2025). [DOI] [PMC free article] [PubMed]

[CR32] 32.Hiramatsu, M. et al. Non-sinus-type dural arteriovenous fistula at the foramen magnum: a review of the literature. J. Neuroendovasc. Ther.19, 2023–0019 (2025). [DOI] [PMC free article] [PubMed]

[CR33] 33.Chen, L. et al. Status of continuous glucose monitoring use and management in tertiary hospitals of China: a cross-sectional study. BMJ Open13, e066801 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR34] 34.Samaniego, E. A. et al. Most promising approaches to improve brain AVM management: ARISE I consensus recommendations. Stroke55, 1449–1463 (2024). [DOI] [PubMed] [Google Scholar]

[CR35] 35.McKeon, A. et al. Pearls & oy-sters: clues for spinal dural arteriovenous fistulae. Neurology76, e10–e12 (2011). [DOI] [PubMed] [Google Scholar]

[CR36] 36.Moosavi, A., Kalapos, P., Church, E. W., Cockroft, K. M. & Thamburaj, K. Localization of spinal dural arteriovenous fistulas from the spatial relationships of perimedullary vessels on standard MRI. J. Neurosurg. Spine40, 389–394 (2024). [DOI] [PubMed] [Google Scholar]

[CR37] 37.Essibayi, M. A., Srinivasan, V. M., Catapano, J. S., Graffeo, C. S. & Lawton, M. T. Spinal dorsal intradural arteriovenous fistulas: natural history, imaging, and management. Neurology101, 524–535 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR38] 38.Kiyosue, H. et al. Angiographic and clinical characteristics of thoracolumbar spinal epidural and dural arteriovenous fistulas. Stroke48, 3215–3222 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR39] 39.Merland, J. J., Riche, M. C. & Chiras, J. Intraspinal extramedullary arteriovenous fistulae draining into the medullary veins. J. Neuroradiol.7, 271–320 (1980). [PubMed] [Google Scholar]

[CR40] 40.Jellema, K., Tijssen, C. C. & van Gijn, J. Spinal dural arteriovenous fistulas: a congestive myelopathy that initially mimics a peripheral nerve disorder. Brain129, 3150–3164 (2006). [DOI] [PubMed] [Google Scholar]

[CR41] 41.Zalewski, N. L. et al. Unique gadolinium enhancement pattern in spinal dural arteriovenous fistulas. JAMA Neurol.75, 1542–1545 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR42] 42.Forbes, G. et al. Complications of spinal cord arteriography: prospective assessment of risk for diagnostic procedures. Radiology169, 479–484 (1988). [DOI] [PubMed] [Google Scholar]

[CR43] 43.Barreras, P. et al. Analysis of 30 spinal angiograms falsely reported as normal in 18 patients with subsequently documented spinal vascular malformations. Am. J. Neuroradiol.38, 1814–1819 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR44] 44.Dice, L. R. Measures of the amount of ecologic association between species. Ecology26, 297–302 (1945). [Google Scholar]

PERMALINK

Deep learning for fast screening and localization of spinal dural arteriovenous fistulas to enhance clinical workflow

Fei Zheng

Xuyang Cao

Jianmin Xu

Nana Wang

Futao Zhang

Lingling Zhang

Kewei Liang

Li Yang

Qi Guo

Yali Wang

Ping Yin

Xuedan Feng

Xuzhu Chen

Nan Hong

Abstract

Introduction

Results

Clinical characteristics

Table 1.

SDAVFdoc model performance

Table 2.

Fig. 1. Evaluation of threshold selection for draining vein cluster identification.

Table 3.

Table 4.

Fig. 2. Performance of the fistula detection model and an illustrative case.

Clinical evaluation

Fig. 3. Quality assessment of automatically segmented draining vein clusters.

Fig. 4. Prospective comparison of processing efficiency between AI and manual methods.

Discussions

Methods

Study design

Fig. 5. Architectural overview of the SDAVFdoc system.

Patient data and CTA images acquisition

Fig. 6. Data flow and evaluation design of the study.

Radiologist’s annotations

SDAVFdoc model development

Clinical evaluation

Statistical analysis

Supplementary information

Acknowledgements

Author contributions

Data availability

Code availability

Competing interests

Footnotes

Contributor Information

Supplementary information

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases