End-to-end deep learning approach for uterine artery-ovarian artery anastomosis detection from digital subtraction angiography sequences.

The requirement of informed consent was waived. All the centers registered and approved the studies labeled as Project Number 2022 (0311). Between July 2018 and July 2023, a cohort of 610 patients underwent UAE, totaling 990 UA angiography sequences and 15 934 DSA images from two hospitals, were included in this study. The dataset from one internal institute was partitioned into a training set (comprising 391 sequences and 7 852 images), a validation set (comprising 197 sequences and 2641 images), and an internal testing set (consisting of 202 sequences and 2736 images). The external testing set was a separate collection of 200 UA angiography sequences (2705 images) from another external hospital. Figure  1 illustrates a flowchart depicting the patient inclusion process for this retrospective study. The Detailed inclusion and exclusion criteria for patients are given in Appendix  S1 . Table  S1 lists the imaging equipments used for the acquisition of DSA images. Flowchart depicting the patient inclusion process for this retrospective study. UAE, uterine artery embolization. UAE procedures were conducted by proficient interventional radiologists with over 15 years of clinical experience. Details of the procedures are given in Appendix  S2 . Our model training necessitated converting DSA sequence frames from DICOM to JPG format and resizing images to 1024*1024 pixels using the nearest neighbor interpolation method, with intensity normalization between 0 and 255. Substandard or irrelevant images were eliminated to refine the dataset quality. Data augmentation techniques, including Mosaic data augmentation, 23 random affine transformations (including rotation, shearing, and scaling), Albumentations (such as blur, grayscale conversion, and histogram equalization), HSV random augmentation, and random image flipping, were implemented using Python (version 3.8.0). The preprocessed JPG image data were used as inputs for training the model. A supervised learning approach (as shown in Figure  2a ) was employed for ground truth annotations. The open‐source program “Labelme” (version 5.3.1) in Python was used to create a polygonal bounding box to outline UA and UOA borders for all 13 805 images. 24 Two interventional radiologists with over a decade of clinical experience undertook the annotation process. The primary annotator marked the data, while the more seasoned radiologist ensured annotation accuracy and refinement. Rectangular bounding boxes identified the UA and UOA with appropriate labels. Overview of UOA‐D's development, evaluation, and clinical application phases. (a) Model development, where the UOA‐D model undergoes training using a supervised learning approach. (b) Model evaluation involves assessing UOA‐D's performance on internal and external test sets alongside a comparative analysis of interventional radiologists’ performance. (c) Clinical translation. DSA, digital subtraction angiography; UA, uterine artery; UOA‐D, uterine artery‐ovarian artery anastomosis detector. UOA‐D incorporates the You‐Only‐Look‐Once version 8 (YOLOv8) algorithm, 25 the latest iteration in the YOLO series, introducing significant advancements over its predecessors. YOLOv8 enhances model efficiency and accuracy through a novel backbone network, an Anchor‐Free detection head, and an updated loss function. This study utilized the YOLOv8 variant, encompassing three principal components: the backbone network (tasked with multi‐scale feature extraction), the neck network (generating feature maps enriched with multi‐scale data), and the head network (responsible for label classification and bounding box predictions), as shown in Figure  S1 . YOLOv8 enhances learning efficiency and reduces computational demands by implementing the backbone network's Cross Stage Partial (CSP) architecture. This architecture innovatively divides the feature map, interlinks the convolution operation's output with the preceding segment, and incorporates the C2f module to amalgamate feature map segmentation with CSP network's (CSPNet's) residual frameworks, enriching the gradient flow information. Furthermore, the neck component utilizes the Feature Pyramid Network (FPN) and Path Aggregation Network (PAN) systems to fuse multi‐scale details comprehensively. The head network, designed to decouple classification and detection tasks, adopts an Anchor‐Free approach for object detection, thereby simplifying the model and enhancing its operational speed. As illustrated in Figure  2c , the application of UOA‐D in clinical settings involved processing angiography sequences of the UA, comprising multiple frames. The model, trained through a fully‐supervised approach, autonomously processed the input data, filtered out targets based on confidence scores exceeding predefined thresholds. The detection results, including UA and UOA, were visually distinguished by different colors and annotated with corresponding labels and confidence scores. The development and training of the model were conducted using the PyTorch 26 framework in Python on an NVIDIA RTX A6000 GPU equipped with 48 GB of memory. Appendix  S3 provides a detailed description of the training setting of the model. As no algorithm is specifically designed for the UA and UOA detection, our UOA‐D model was benchmarked against established object detection algorithms, including Faster Region‐Based Convolutional Neural Network (Faster‐RCNN) 27 and Detection Transformer (DETR), 28 on the internal testing set to assess detection performance. The training for these comparisons was carried out over 100 epochs using the training set, with training parameters detailed in Table  S2 . Performance metrics such as the area under the Precision‐Recall (PR) curve, the area under the receiver's operating characteristic (ROC) curve, average precision at Intersection over Union (IoU) threshold 0.5 (AP_0.5), and average precision across a range of IoU thresholds from 0.5 to 0.95 (AP_0.5:0.95, namely AP) were utilized for evaluation. The assessment involved three junior interventional radiologists with >5 years of UAE procedure experience and three senior radiologists with >15 years of experience. They evaluated the UA angiography sequences from both testing sets and annotated UOA on a specific frame of the DSA sequence. The total time taken by the interventional radiologists to review and annotate all sequences in the test set was recorded, and the average time per sequence was calculated. UOA detection of UOA‐D is based on DSA sequence. To balance sensitivity and specificity, the confidence threshold was set to 0.5, determined by analyzing the ROC curve to maximize the Youden's J statistic. For each DSA sequence, as long as a UOA is accurately detected in at least one frame, it can be considered as true positive. The detection time for UOA‐D was defined as the interval from the start of the detection process to the generation of the annotated result sequences. Similarly, the average time per sequence was calculated. Sensitivity, specificity, accuracy, and F1 score were the metrics used to compare the performance of interventional radiologists and UOA‐D. One week after the initial evaluations, junior and senior interventional radiologists reevaluated the UOA with detection results of UOA‐D, presented separately for side‐by‐side comparison. They were permitted to adjust their initial findings or maintain them without making any changes. Moreover, Fleiss's kappa coefficient was calculated to measure the level of agreement between the two radiologist groups under both assisted and unassisted scenarios. For continuous variables, such as patient age, normality and homogeneity tests were first conducted. Depending on the outcomes, either the analysis of variance (ANOVA) or the Mann–Whitney U test was applied. Categorical variables, such as uterine artery location and UOA presence, were analyzed using Pearson's chi‐square test. The performance of the detection models was evaluated using metrics including area under the ROC curve (AUC), AP_0.5, AP, sensitivity, specificity, accuracy, and F1 score. Statistical analyses were performed using R (version 2023.03.1) and Python software (version 3.8.0), with a P‐value <0.05 denoting statistical significance.

This investigation included 610 female patients from July 2018 to July 2023, dividing the dataset into training, validation, internal testing, and external testing sets, with 262, 115, 121, and 112 patients, respectively (corresponding to 391, 197, 202, and 200 uterine artery angiography DSA sequences for each set). The mean age of patients in these subsets was 43.5 ± 5.5 (standard deviation, SD), 42.6 ± 8.4 (SD), 42.0 ± 9.1 (SD), and 41.5 ± 6.7 (SD), respectively, indicating no significant age difference across the groups ( p = 0.945). Comprehensive clinical characteristics of the patients in each subset are summarized in Table  1 . The distribution concerning the uterine artery's side and the occurrence ratio of UOA was comparable across the datasets, with no significant distribution variations observed among the subsets, as depicted in Table  1 and Figure  S2 . Baseline characteristics of patient datasets. Note : Data were presented as number of patients (%). Abbreviations: SD, standard deviation; UA, uterine artery; UOA, uterine artery‐ovarian artery anastomosis. UOA‐D outperformed the other two deep learning algorithms (Faster‐RCNN and DETR) in detecting UA and UOA. As shown in Figure  3a , in the task of UA detection within the internal testing set, UOA‐D attained a higher area under the PR curve than Faster‐RCNN and DETR (0.9833, 0.9509, 0.7678, respectively). Similarly, for the UOA detection task in the internal testing set, UOA‐D recorded a higher area under the PR curve than that achieved by Faster‐RCNN and DETR (0.7485, 0.6775, 0.4534, respectively), as shown in Figure  3b . Baseline comparative evaluation of UOA‐D, Faster‐RCNN, and DETR Models. (a) PR curves comparing the models for detection in the internal testing set; (b) PR curves for UOA detection of the models in the internal testing set; (c–f) Comparative sensitivity, specificity, accuracy, and F1 score across the three models for identifying UOA‐positive cases in the internal testing set; (g) Visualization of model performance in detecting UA and UOA in the internal testing set. Red box and red text indicate UA and confidence scores; yellow box and yellow text indicate UOA and confidence scores. Patient "1", a 39‐year‐old woman with uterine fibroids post‐UAE, is shown in the left uterine artery angiography image; Patient 2, a 45‐year‐old woman with uterine fibroids post‐UAE, is depicted in the right uterine artery angiography image. DETR, Detection Transformer; RCNN, Region‐Based Convolutional Neural Network; UAE, uterine artery embolization. In sequence‐inference performance within the internal testing set, UOA‐D surpassed both Faster‐RCNN and DETR across sensitivity, specificity, accuracy, and F1 score metrics, as shown in Figure  3c–f . Specifically, sensitivity rates for UOA‐D, Faster‐RCNN, and DETR were 0.975, 0.934, and 0.791, respectively; specificity rates were 0.942, 0.716, and 0.815, respectively; accuracy rates were 0.967, 0.853, and 0.807, respectively; and F1 scores of 0.973, 0.883, and 0.829, respectively, in the internal testing set. Figure  3g presents the visualization results of the three models in detecting UA and UOA. For Patient 1, who has a UOA, the DETR model identified three potential regions but suffered from inaccurate bounding box selection. The Faster‐RCNN model failed to fully encompass the UOA within its bounding box. In contrast, UOA‐D provided the most accurate detection, closely matching the gold standard. For Patient 2, the UOA was relatively small and not well visualized. Neither DETR nor Faster‐RCNN was able to detect it. However, UOA‐D successfully identified the UOA. These results clearly demonstrate that UOA‐D significantly outperformed other object detection models. Initially, evaluations were conducted on internal and external testing sets using the DSA image dataset. For UA detection, UOA‐D recorded an AUC of 0.981 in the internal testing set (Figure  4a ) and 0.971 in the external testing set (Figure  4b ). As shown in Figure  4c , in the internal testing set, AP and AP_0.5 were 0.743 and 0.982, respectively. In the external testing set, AP and AP_0.5 were 0.736 and 0.980, respectively. Performance of UOA‐D in detecting UA and UOA across testing sets. (a, b) ROC curves of UOA‐D for UA detection in the internal and external testing sets, respectively; (c) AP and AP_0.5 values for UA detection in the internal and external testing sets; (d, e) ROC curves for UOA detection by UOA‐D in the internal and external testing sets, respectively; (f) AP and AP_0.5 values for UOA detection in both testing sets; (g, h) Confusion matrices showcasing UOA‐D's accuracy in distinguishing UA and UOA in both testing sets; (i–l) Sensitivity, specificity, accuracy, and F1 score of UOA‐D in identifying UOA‐positive cases in the internal and external testing sets. AP, average precision at IoU threshold 0.5:0.95; AP_0.5, average precision at IoU threshold 0.5; ROC, receiver's operating characteristic; UA, uterine artery; UOA, uterine artery‐ovarian artery anastomosis. However, the performance of UOA‐D detecting UOA, a considerably smaller target, was slightly inferior. UOA‐D attained an AUC of 0.831 in the internal testing set (Figure  4d ) and 0.829 in the external testing set (Figure  4e ). As shown in Figure  4f , in the internal testing set, AP and AP_0.5 were 0.541 and 0.773, respectively. In the external testing set, AP and AP_0.5 were 0.509 and 0.768, respectively. Detection performance on validation set is illustrated in Figure  S3 . The distinction between UA and UOA by UOA‐D was further evaluated through confusion matrices in both the internal (Figure  4g ) and external testing sets (Figure  4h ), demonstrating UOA‐D's precision in accurately identifying and differentiating between UA and UOA. Moreover, to assess UOA‐D's clinical functionality, tests were performed using DSA sequences in both testing sets. In the internal testing set, UOA‐D exhibited sensitivity, specificity, accuracy, and F1 score of 0.978, 0.951, 0.961, and 0.978, respectively, for detecting UOA‐positive sequences. Similarly, in the external testing set, UOA‐D achieved sensitivity, specificity, accuracy, and F1 score of 0.969, 0.945, 0.959, and 0.962, respectively (as shown in Figure  4i–l ). Figure  5 illustrates the sequence inference outcomes for UOA‐D, featuring Patients “1” and “2” from internal set, and Patients “3” and “4” from the external testing sets, underscoring its adeptness at accurately detecting targets of varying sizes. Detailed detection outcomes for four cases across the internal (Patients "1" and "2") and external testing sets (Patients "3" and "4"). Each row represents a different patient, with columns progressing from left to right to illustrate the dynamic contrast‐enhanced imaging over time. Red box and red text indicate UA and confidence scores; yellow box and yellow text indicate UOA and confidence scores. Patient "1", a 50‐year‐old woman treated for uterine fibroids through UAE, is represented in the right uterine artery angiography, revealing a slender UOA. Patient "2", a 34‐year‐old woman treated for uterine fibroids via UAE, is represented in the left uterine artery angiography indicating a substantial presence of UOA. Patient "3", aged 48, underwent UAE for uterine fibroids and concurrent bleeding; her left uterine artery angiography distinctly shows UOA. Lastly, Patient "4", a 40‐year‐old woman who received UAE for uterine fibroids, is depicted with a relatively small UOA in the left uterine artery angiographic image. UA, uterine artery; UAE, uterine artery embolization; UOA, uterine artery‐ovarian artery anastomosis. As detailed in Table  2 , senior interventional radiologists achieved higher sensitivity, specificity, and accuracy in detecting UOA, while spending less time, compared with junior interventional radiologists (Table  S3 ). UOA‐D demonstrated superior sensitivity, specificity, and accuracy in detecting UOA compared to junior interventional radiologists across internal and external testing sets while offering results comparable to those of senior interventional radiologists. A notable advantage of UOA‐D is its significantly reduced evaluation time per case, which is markedly less than that required by both junior and senior interventional radiologists (Internal testing set: 3.0, 8.5, 7.4 s, respectively; external testing set: 2.9, 8.8, 7.8 s, respectively). Performance in detecting UOA: A comparative analysis of UOA‐D and interventional radiologists (with and without UOA‐D assistance). Note : Data are presented as mean (95% CI). Abbreviations: CI, confidence interval; UOA, uterine artery‐ovarian artery anastomosis; UOA‐D, UOA detector. Table  2 illustrates that the introduction of UOA‐D has markedly enhanced the performance of junior interventional radiologists in UOA detection, with sensitivity, specificity, and accuracy increases of 9.5%, 6.7%, and 8.4% in the internal testing set and 9.8%, 12.7%, and 11.3% in the external testing set, respectively. As indicated in Table  S3 , their detection capabilities are comparable to those of senior interventional radiologists. Furthermore, senior interventional radiologists, leveraging UOA‐D, achieved the highest accuracy across both datasets. They utilized UOA‐D's detection outcomes to identify false negative instances that had been overlooked initially (Figure  6a ). Incorporating UOA‐D accelerated the evaluation process for radiologists of all experience levels. Specifically, detection times for junior and senior interventional radiologists decreased by 49.4% and 44.6% in the internal testing set, and by 50.0% and 52.6% in the external testing set, respectively. Consistency analyses depicted in Figures  6b and c reveal that with the assistance of UOA‐D, both junior and senior radiologists significantly improved their consistency in detection, attaining a high degree of agreement. Enhancement in UOA detection by Interventional radiologists with UOA‐D support. (a) Examples of cases in which UOA was missed by the interventional radiologists but was detected by UOA‐D. The ground truth image is a subsequent frame from the same patient, in which the UOA has become distinctly visible. Red box and red text indicate UA and confidence scores; yellow box and yellow text indicate UOA and confidence scores. (b, c) The agreement of three junior radiologists, three senior radiologists, and all six radiologists before and after utilizing UOA‐D in internal (b) and external (c) testing sets.

The proposed deep learning model UOA‐D facilitated the rapid identification of UOA. By enabling the detection of UOA during UAE procedures and providing detailed insights into the vascular network and its dimensions, UOA‐D aids interventional radiologists in adjusting the dosage, injection pressure, and rate of embolic material administration. Specifically, the early identification of UOA allows for a reduction in the dosage of embolic material, minimizing the risk of ovarian ischemia and other complications. It also prompts a decrease in injection pressure and a slowdown in the injection rate, ensuring that the embolic material is primarily distributed within the uterine artery and reducing the likelihood of it entering the ovarian artery through the anastomosis. This precision reduces the risk of unintentional embolization and enhances the concentration of embolic materials in the targeted uterine arteries. 29 , 30 Consequently, the likelihood of procedural complications is lowered, thereby enhancing the safety of UAE interventions. The challenge of detecting UOA, given its diminutive size and atypical presentation, frequently exceeds the capabilities of traditional radiology imaging techniques. A previous investigation 10 employed thin‐slice computed tomography (CT) scans combined with vascular casting to identify UOA, an approach that incurs considerable expense. Moreover, undergoing additional examinations solely for preoperative UOA detection is impractical. In contrast, our UOA‐D model offers a swift and precise solution for identifying UOA and UA from intraoperative DSA sequences, providing timely assistance to interventional radiologists. The results indicated that the UOA‐D exhibited commendable performance in detecting UA, paralleling or surpassing other models. However, its proficiency in identifying UOA surpassed that of the compared models, an advantage attributed to YOLOv8's comprehensive image‐processing abilities. This model processes images holistically to directly predict bounding boxes and their corresponding confidence levels, 25 , 31 which enhances detection efficiency and accuracy. In contrast, the region‐based approaches of Faster‐RCNN and the transformer‐based architecture of DETR may lead to suboptimal performance in detecting small and intricate structures such as UOA. 32 Specifically, while DETR reliably identifies the larger UA, it struggles to detect relatively minor UOA with similar confidence, highlighting the distinct advantages of UOA‐D in such applications. The model's performance was evaluated against interventional radiologists to assess the practical application in detecting UOA. Furthermore, the detection capabilities of interventional radiologists were examined, both with and without UOA‐D assistance. The results indicate that UOA‐D significantly enhanced the speed and accuracy with which radiologists of varying experience levels can identify UOA, promoting a high degree of consistency. These results suggest that UOA‐D could be instrumental in overcoming real‐world clinical challenges, offering substantial assistance, and enhancing the proficiency of interventional radiologists, particularly those with less experience. This study has several limitations. First, there are a few false positives and false negatives in some special cases. The model tends to overlook insignificant UOA when the vessels are small or the contrast agent fails to fill adequately. It may also misjudge the presence of UOA when image quality is poor or when surrounding vessels cause occlusion. In the future, we plan to introduce attention mechanisms in algorithm development to focus more on the location and anatomical features of UOA, thereby improving its detection capabilities for small targets and low‐contrast areas. In clinical applications, optimizing the injection parameters of the contrast agent (such as dosage, injection pressure, and rate) to improve the clarity of UOA angiography and reduce background vessel interference is also a potential approach. Moreover, the limited diversity of the data sources is another limitation of our study. Although the research incorporated external testing, the data were limited in terms of variability in imaging equipment, procedural differences in UAE, and diversity of patient vascular anatomy, all of which could potentially influence detection outcomes. This underscores the necessity for validation through large‐scale prospective clinical trials.

In conclusion, this research has effectively demonstrated the feasibility and accuracy of the deep learning‐based object detection model, UOA‐D, for identifying UOA during interventional procedures. UOA‐D's encouraging performance underscores its significant potential for clinical application. Comprehensive, prospective, and multicenter real‐world clinical trials could offer more robust evidence for the clinical application of UOA‐D.

Uterine fibroids represent the predominant type of tumor among women of reproductive age, with a cumulative incidence reaching approximately 70%. 1 , 2 , 3 These tumors can cause a range of distressing symptoms, including heavy menstrual bleeding, abdominal pain, and a sensation of abdominal pressure, adversely impacting the quality of life of nearly half of the affected individuals. 4 In 1995, uterine artery embolization (UAE) was introduced as a minimally invasive intervention for treating uterine fibroids. 5 Since 2008, the American College of Obstetricians and Gynecologists (ACOG) has endorsed the minimally invasive UAE for women desiring to preserve their uterus while suffering from fibroids. 6 Furthermore, UAE has been successfully employed in addressing other gynecological conditions, such as adenomyosis, 7 uterine vascular malformations, 8 and postpartum hemorrhage. 9 However, specific vascular anatomical structures, such as the Uterine Artery‐Ovarian Artery Anastomosis (UOA), introduce additional challenges to UAE's safety and efficacy. UOA denotes a vascular link between the uterine and ovarian arteries within the pelvic cavity. 10 UOA facilitates a collateral circulation route between the arteries, providing an alternative blood flow pathway in blockage or compromised supply events. However, this also poses some challenges for UAE treatment. Firstly, embolic agents injected through the uterine artery may get shunted to the ovarian artery, leading to unintended embolization and reduced ovarian blood supply. This shunting may adversely affect ovarian development and endocrine function, potentially resulting in decreased ovarian reserve, infertility, or premature menopause. 11 , 12 Secondly, diverting embolic agents into the ovarian artery may lead to incomplete necrosis of uterine fibroids, 12 , 13 , 14 which could necessitate additional embolization procedures. 15 Kim et al. have demonstrated that compared with patients without UOA visualization during the procedure, those with UOA visualization have a higher proportion of requiring repeat embolization after UAE. 15 , 16 What's more, Sheikh GT et al. indicated that patients with Type Ib and Type III anastomoses were at risk of experiencing ovarian failure following UAE. Therefore, it was recommended to apply protective coiling to mitigate this risk and prevent ovarian failure in such cases, 17 which was in line with the conclusions drawn by Salazar et al. 18 Consequently, interventional radiologists must pay greater attention and exercise caution when UOA is present. However, identifying UOA is challenging due to their small size, anatomical variability, 19 and the transient nature of contrast agent flow, which complicates detection with the naked eye. These factors contribute to the clinical under‐recognition of UOA. Recent advances in deep learning have demonstrated significant achievements in diagnosing vascular‐related conditions, 20 attributed to their robust feature extraction capabilities. While deep learning has facilitated breakthroughs in detecting conditions such as intracranial vascular perforations 21 and carotid artery stenosis 22 from digital subtraction angiography (DSA) sequences, research on employing these algorithms for UOA detection remains unexplored. This study developed a deep learning model, the UOA Detector (UOA‐D), capable of real‐time detection and localization of UOA. Recognizing the spatial and temporal association between the uterine artery (UA) and UOA in DSA sequences, our approach included simultaneous UA detection to enhance UOA identification accuracy and minimize false positives. The performance of this model was further assessed against that of interventional radiologists, underscoring its clinical applicability and potential to improve diagnostic accuracy and procedural safety in the UAE.

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.

Supporting Information