RVO-ME: A Dual-Task OCT Dataset for Segmentation and Detection of Macular Lesions in Retinal Vein Occlusion

Abstract

Retinal vein occlusion (RVO) is one of the most common vision-threatening retinal diseases, with macular edema (ME) as its primary complication. Optical coherence tomography (OCT), a non-invasive imaging modality, enables detailed visualization of retinal structures and fluid distribution, thus supporting accurate diagnosis, treatment monitoring, and clinical assessment of RVO-related conditions. However, the development of automated algorithms for RVO-ME analysis has been hindered by the lack of high-quality, manually segmented datasets. To address this limitation, we constructed a manually annotated RVO-ME dataset comprising 3,012 OCT B-scans from 146 eyes of 130 patients. For each image, we provide segmentation labels for four key retinal features (subretinal fluid, intraretinal fluid, the ellipsoid zone, and the external limiting membrane), along with point annotations to facilitate the detection of highly reflective foci. This dataset provides a valuable benchmark for assessing the performance of segmentation algorithms and facilitates the advancement of artificial intelligence models for RVO-related disease analysis.

Background & Summary

Retinal Vein Occlusion (RVO) is the second most common retinal vascular disease after diabetic retinopathy and represents a significant cause of vision loss worldwide^1,2. According to global epidemiological studies, the prevalence of RVO in the population aged 30-89 years is 0.77%, indicating that approximately 28 million people worldwide are affected by this condition, with the potential for a decrease in visual function³. Macular edema (ME) is a common complication of RVO and the primary cause of vision loss in these patients. Therefore, early detection and timely treatment of ME are critical to improving visual outcomes^4–8. Research has shown that vascular endothelial growth factor (VEGF) is a crucial cytokine that mediates increased vascular permeability, subsequently leading to ME^9,10. Anti-VEGF therapies have become the standard treatment approach, demonstrating significant efficacy in improving visual acuity (VA) and reducing ME^11–15.

Optical coherence tomography (OCT), introduced by Huang et al. in 1991¹⁶, is a noninvasive imaging technology that has become widely used in the diagnosis and treatment of ocular diseases due to its advantages of simplicity, painlessness, and non-contact nature¹⁷. Its penetration depth is not limited by transparent ocular media, enabling a clear visualization of retinal structures and macular fluid, thus providing physicians with intuitive information on the retinal layers and associated features. This helps in the early diagnosis and evaluation of retinal pathologies, particularly the presence of ME. OCT allows for the quantitative assessment of retinal thickening and provides detailed information on lesion characteristics such as subretinal fluid (SRF), intraretinal fluid (IRF), ellipsoid zone (EZ), external limiting membrane (ELM), and highly reflective foci (HF). Through qualitative and quantitative analysis of retinal lesions, physicians can more accurately assess disease activity and formulate personalized treatment plans for RVO patients. OCT provides key structural features, such as the presence and extent of SRF and IRF, disruption of the EZ, and central retinal thickness (CRT), which serve as valuable biomarkers for evaluating response to treatment and predicting the visual prognosis. Furthermore, the researchers aimed to explore certain characteristics of OCT that are associated with a poor response to anti-VEGF therapy. Identifying patients with these characteristics in advance can help predict suboptimal treatment results and reduce the frequency of unnecessary intravitreal injections. Therefore, OCT technology has particular clinical value for patients who exhibit low responsiveness to standard therapies, allowing for more informed treatment decisions and reducing both psychological and economic burdens.

Accurate segmentation of morphological patterns is crucial for quantifying the severity of Retinal Vein Occlusion Macular Edema (RVO-ME), as it assists in providing imaging biomarkers for predicting treatment outcomes. With the rapid development of artificial intelligence (AI) in the medical field¹⁸, various deep learning models based on AI have been developed. These models focus primarily on the diagnosis and prognostic prediction of RVO, including diagnostic and classification models based on color fundus photography^19–22 and prognostic prediction models based on OCT images^23,24, all of which demonstrate superior specificity and sensitivity. However, the application of segmentation models based on OCT images in clinical practice remains limited. This limitation may arise from the dependency of AI-based retinal pathology segmentation models on large-scale datasets of annotated retinal images. The training of these models requires large datasets with ground truth segmentation²⁵. Manual segmentation performed by experienced ophthalmologists is the most reliable method for obtaining such gold-standard datasets. However, this process is highly specialized and time-consuming. Moreover, the scarcity of public datasets further hinders the advancement of these models.

Previous studies on OCT image analysis of RVO-ME have primarily focused on the segmentation of retinal fluid. For example, the Retouch dataset provides OCT images of retinal fluid segmentation in conditions such as age-related macular degeneration (AMD), diabetic macular edema (DME), and RVO²⁶. However, OCT imaging reveals that the pathological changes in RVO are diverse, with fluid accumulation being only one of the manifestations. Research by Sen²⁷et al. demonstrated that the integrity of the EZ is a significant predictor of visual improvement in RVO-ME patients after treatment. Subsequently, studies by Segal et al.²⁸ and Cunha Ferreira et al.²⁹ further confirmed that the presence of the ELM is also a key biomarker influencing RVO prognosis. It is noteworthy that HF, as an important imaging biomarker, may reflect the pathophysiological mechanisms of ME caused by various etiologies³⁰, due to its unique characteristics.

The study confirms that the integrity of the EZ, the presence of IRF, SRF, ELM, and HF are critical biomarkers influencing vision loss and prognosis in patients with RVO^31,32. Based on this, the research team, consisting of experienced ophthalmologists, meticulously manually annotated the aforementioned features in OCT images. It is particularly noteworthy that, due to the small size of HF, its annotation process is prone to observer bias. Therefore, during the construction of this dataset, targeted validation experiments were specifically conducted on HF to enhance the accuracy and reliability of its annotation. Ultimately, a high-quality dataset was successfully established, incorporating these five RVO-ME lesion features. The creation of this dataset provides a solid data foundation for the development of artificial intelligence models targeting RVO diseases and will significantly promote the exploration of AI technologies in the clinical diagnosis and treatment of RVO.

Methods

Data collection

This study was reviewed and approved by the Ethics Committee of the Second Affiliated Hospital of Nanchang University (approval number: O-Medical Research Ethics [25] No. 132) and strictly adhered to the ethical principles of the Declaration of Helsinki. All participants provided their informed consent in writing, acknowledging that their images may be used for future clinical research and the dissemination of public data sets. For data protection, all personally identifiable information was de-identified and image data were stored under anonymized codes. Access was restricted to authorized research personnel in a secure environment. This study was conducted from June 2019 to October 2024 and included 130 patients diagnosed with retinal vein occlusion macular edema (RVO-ME), involving 146 eyes and a total of 3,012 OCT B-scan images. In our previous work, we constructed a publicly available dataset for wet AMD lesions and key structures in the macular area using a similar data collection process³³. We adopted essentially the same pipeline for patient recruitment, data collection, and image annotation.

Specifically, the diagnosis of patients was made by three primary ophthalmologists, each with a minimum of three years of clinical experience, and one senior ophthalmologist with at least eight years of clinical experience. These patients were then incorporated into the dataset. The exclusion criteria included: (1) patients who had undergone vitrectomy or had intraocular implants in the vitreous cavity; (2) patients with a prior diagnosis of other vitreoretinal diseases; (3) patients with poor imaging quality, defined as a signal strength lower than 5/10 when using the Cirrus HD-OCT 5000 scanner, or incomplete horizontal B-scan images with missing portions or incomplete tissue structure representation.

All OCT images were captured using the Zeiss Cirrus HD-OCT 5000 device (Zeiss, Germany) by experienced ophthalmologists. During the imaging process, patients were seated with their gaze aligned at a 90-degree angle to the integrated video monitor. The chin was placed on a support device, and the head was stabilized by a supporting frame to ensure precise alignment of the eyes with the fovea. OCT B-scan images were obtained both before and after treatment, allowing for a more comprehensive capture of the clinical features of RVO-ME patients. This approach facilitated subsequent structural segmentation analysis. Figure 1 provides an overview of the data collection and annotation process.

Fig. 1.

The workflow for establishing the RVO-ME dataset (a) Data collection process. 130 patients were diagnosed with RVO, and 3012 OCT images were collected at the Department of Ophthalmology of The Second Affiliated Hospital of Nanchang University. (b) Annotation process. Junior ophthalmologists first performed manual segmentation of IRF, SRF, EZ, and ELM structures and annotated HF locations; subsequently, a senior retinal specialist reviewed each initially labeled image until every case passed the expert’s final approval.

The intra-consistency training

Before launching the large-scale lesion segmentation project, we designed a dedicated training round for junior ophthalmologists to ensure that all participants shared a uniform, pixel-level understanding of RVO-related findings on OCT B-scans. Three trainees were asked to annotate the same set of ten representative B-scans independently, on different days, and without mutual communication. Pixel-wise agreement was then quantified for four key structures: SRF, IRF, ELM, and EZ. As shown in Fig. 2, SRF achieved an IoU of 0.953 and a Dice of 0.974; IRF reached 0.979 and 0.989; ELM 0.943 and 0.969; EZ 0.949 and 0.969.

Fig. 2.

The Intra-annotator agreement evaluation. (a) The average IoU was from 0.9054 to 0.9634. (b) The average Dice coefficient was from 0.9378 to 0.9834.

Although overall concordance was high, data analysis revealed noticeably lower DSC (Dice Similarity Coefficient,DSC) and IoU (Intersection of Union, IoU) values for ELM and EZ. This discrepancy was traced to the fact that both layers occupy only a small pixel fraction in the RVO data set and exhibit variable, often blurred, boundaries. Consequently, early-stage annotators lacked sufficient “shared visual memory” for these classes, leading to greater inter-rater dispersion. This internal-consistency check not only quantified cognitive differences across experience levels but also identified two focus areas for subsequent training and quality control: (1) expanding the teaching gallery for under-represented lesions, and (2) introducing a real-time consensus mechanism in which senior experts provide immediate feedback on critical slices. These measures established a more robust, quantifiable baseline for future collaborative annotation efforts.

Image Annotation Process

The annotation team for the patients’ OCT B-scan images consisted of three junior ophthalmologists and one senior ophthalmologist. Initially, the three junior ophthalmologists manually annotated and segmented the acquired raw OCT images using the LabelMe software. Subsequently, the senior ophthalmologist conducted a quality review and refined the annotations. Before beginning the annotation, the team members engaged in thorough discussions regarding the characteristics of the lesions and reached a consensus. All members collectively learned the usage and operational procedures of the LabelMe software to ensure a unified understanding of its functionalities and workflow. Subsequently, the three junior ophthalmologists individually used the LabelMe software to manually annotate five primary lesion characteristics: IRF, SRF, EZ, ELM, and HF.

The annotation requirements are outlined as follows: (1)IRF: refers to a hyporeflective dark area located within the retinal neuroepithelial layer. (2)SRF: a hyporeflective dark area situated between the retinal neuroepithelial layer and the retinal pigment epithelium (RPE). It is essential to differentiate SRF from certain IRF regions near the RPE to avoid misinterpretation. (3)EZ: a normal anatomical structure rich in mitochondria, marking the junction between the inner and outer segments of photoreceptors. It is located between the external limiting membrane and the RPE, appearing as a hyperreflective signal band. The rupture of the EZ is often irregular, and in some cases, it may rupture centrally, potentially forming multiple discontinuous bands. Therefore, when annotating, it is necessary to label from left to right to ensure no segments are missed. If the EZ cannot be identified or annotated due to image quality issues, this region should be left blank. (4)ELM: a thin, mesh-like structure composed of the junctions between adjacent photoreceptors and Müller cells. When annotating, care should be taken to label from left to right to avoid omissions. (5)HF: Characterized by discrete, punctate, high-reflective signals in various retinal layers (especially from the inner nuclear layer to the outer plexiform layer), typically with a diameter of less than 30 microns. These signals form a distinct contrast with the surrounding tissue. The underlying pathology involves activated microglial cells or lipoprotein exudates. Given the varying characteristics of different structures in OCT, SRF and IRF were segmented and labeled using polygons to delineate the lesion structures. ELM and EZ were outlined in the form of line segments to indicate their locations. Since HF typically appears as bright spots occupying only a few pixels in OCT images, they were labeled in the form of points.

This investigation prioritized the preservation of unaltered tissue morphology in OCT B-scans by strictly prohibiting pixel manipulation during annotation. This protocol prevented artificial disruption of anatomical continuity, effectively suppressing false-positive labeling artifacts. Representative B-scan slices (20-60 per case) displaying characteristic pathological features were systematically selected, with the spatial distribution across macular lesions maximized to capture structural heterogeneity. Pixel-accurate delineations were executed exclusively on these slices. Crucially, scans dominated by normal retinal architecture, requiring negligible annotation, were deliberately omitted from the analytical cohort.

All initially annotated images completed by primary ophthalmologists were evaluated by expert ophthalmologists, who made corresponding decisions based on the assessment. The specific process is shown in Fig. 1b. After the annotations were completed by three primary doctors, expert ophthalmologists performed a quality review and rating of the annotated OCT images. The images were classified into three grades: A, B, and C. Grade A represents excellent quality, passing the quality check with no need for modifications; Grade B represents acceptable quality, requiring modification of the OCT image annotations under the guidance of an expert; Grade C represents unacceptable quality, necessitating re-annotation by the initially responsible physician, followed by a re-evaluation by the expert. The entire annotation process was strictly followed by the aforementioned procedure. This grading approach ensured both the efficiency of the annotations and, through the expert final review, the accuracy of the medical image annotations, thereby laying a reliable foundation for the subsequent dataset.

It is worth mentioning that during the annotation process, the annotating physicians also annotated the epiretinal membrane (ERM). The annotation requirements are as follows: ERM: Located at the vitreoretinal interface of the macular region, it is a highly reflective band. The membrane is formed by the abnormal proliferation, migration, and deposition of retinal pigment epithelium cells, glial cells (e.g., Müller cells), fibroblasts, inflammatory cells, and extracellular matrix components such as collagen and fibrin secreted by these cells. Although ERM was annotated, its overall prevalence in the collected OCT B-scans was relatively low. Quantitative analysis revealed that ERM was present in only 23 cases. Given this relatively low prevalence and the associated insufficient sample size for robust statistical analysis and effective deep learning model training, ERM was excluded from the final target feature set. This decision was made to maintain dataset balance and ensure adequate representation for all included pathological structures. Importantly, while ERM annotations are not included in the final mask images, the original LabelMe JSON files retain these annotations, providing flexibility for future research initiatives that may specifically focus on ERM detection or incorporate larger ERM datasets.

Data Records

The RVO-ME dataset³⁴ is available at Figshare under 10.6084/m9.figshare.29804435.v1. And the dataset has undergone preliminary usability validation by peer-review experts.

All images and corresponding annotations are stored under a single root directory that contains three subfolders and one Excel file. The Excel file, named ‘Demographics of the Participants’, lists row-by-row patient ID, Eye ID, Eye category, Age (years), Gender, and other demographic variables for subsequent stratification and statistical analysis. The Image_Seg folder houses the data required for segmentation tasks: its images subdirectory stores the original OCT-B-scans, and the masks subdirectory contains the pixel-wise labels. The train.txt and test.txt files were partitioned at the patient level, with approximately 80% of the patients (104 individuals) assigned to the training set and the remaining 20%(26 individuals) allocated to an independent test set. This patient-level division ensures that no subject appears in both datasets, thereby preventing data leakage and preserving the integrity of the evaluation. Within each mask, the background is encoded as 0, SRF as 1, IRF as 2, ELM as 3, and EZ as 4.

Among them, SRF and IRF were annotated as polygonal regions encompassing the entire area marked by clinicians. In contrast, ELM and EZ represent structural boundaries that appear as distinct dividing lines in retinal images. Therefore, these structures were annotated using linear segments rather than enclosed regions. To facilitate downstream deep learning applications, these linear annotations were rendered as 2-pixel-wide lines in the corresponding mask images. This minimal width was carefully chosen to provide sufficient spatial information for convolutional neural networks to effectively learn these anatomical boundaries, while ensuring minimal encroachment on adjacent retinal layers. The narrow profile prevents overlap with neighboring structures, thereby preserving the distinct morphological characteristics of individual retinal interfaces. Notably, as we provide the original JSON annotation files, users have the flexibility to adjust the line width according to their specific requirements.

Since HF appears as punctate lesions that are too small for reliable segmentation, they are not labeled in the masks. The RVO_HF_Detection folder is dedicated to detection tasks: every HF focus identified by expert readers is circumscribed with a 10  × 10-pixel bounding box. The RVO_Lesion_LabelMe folder retains the full set of lesion contours drawn with the LabelMe software; these polygons are saved as raw JSON files to facilitate later auditing and visualization. All images follow the uniform naming convention ‘n_x.png’, where n denotes the Eye ID, and x indicates the sequential image number within that eye, ensuring complete traceability and non-redundancy throughout the study workflow. Figure 3 illustrates the publicly released root directory structure of the RVO-ME dataset.

Fig. 3.

Data records root structure of RVO-ME dataset.

The dataset comprises 3,012 OCT B-scan images, each accompanied by its corresponding lesion segmentation mask, all at a resolution of 570  × 380 pixels; the main characteristics are summarized in Table 1. The scans were acquired from 146 eyes of 130 participants—58 women and 72 men—with a mean age of 67.1  ± 9.3 years; all participants were of Asian descent.

Table 1.

Data statistics of RVO-ME.

Item	Value
Total number of participants	130
Number of eyes	146
Number of OCT B-scans	3,012
Age, year (mean  ± s.d.)	67.1 (9.3)
Female (%)	58 (44.6%)

Technical Validation

To verify that this dataset is a viable basis for RVO evaluation on OCT B-scans, we conducted two parallel experiments: supervised structural segmentation and HF lesion detection. All scans were randomly split 8:2 into training and test sets.

For segmentation, the Dice coefficient served as the baseline metric. We compared U-Net³⁵, U-Net++³⁶, Att-UNet³⁷, ResUNet³⁸, and FNeXter³⁹. Table 2 reports quantitative results for each structure and model; Fig. 4 shows representative B-scans, expert labels, and model predictions. U-Net++ performed best, achieving a mean IoU of 52.99 and a mean Dice of 68.52. Overall performance was stable, yet thin structures, such as ELM and EZ, were segmented less accurately than SRF and IRF. Our segmentation pipeline was built upon an open-source toolkit⁴⁰, and all task-specific adaptations are publicly available in our GitHub repository.

Table 2.

Baseline image segmentation model results.

Metrics	Type	Method
		U-Net35	U-Net++36	Att-UNet37	ResUNet38	FNeXter39
IoU	SRF	68.57	69.01	70.65	69.29	69.10
	IRF	58.52	58.00	56.10	56.89	56.83
	ELM	37.67	38.33	37.49	37.82	38.56
	EZ	43.81	46.62	45.48	45.91	44.79
	Mean	52.14	52.99	52.43	52.48	52.32
Dice	SRF	81.36	81.67	82.80	81.86	81.73
	IRF	73.83	73.42	71.88	72.53	72.47
	ELM	54.72	55.42	54.53	54.89	55.66
	EZ	60.93	63.59	62.52	62.92	61.87
	Mean	67.71	68.52	67.93	68.05	67.93

Fig. 4.

Comparison of visual segmentation results. Red, blue, green, yellow, and pink colors represent SRF, IRF, ELM, and EZ, respectively.

Since HF appears as tiny hyper-reflective pixels, we designed a complementary detection experiment. The same 8:2 split was used, with SSD⁴¹ and Faster-RCNN⁴² as the detectors. Point annotations in LabelMe were converted into 10  × 10-pixel bounding boxes centered on each labeled point. Table 3 summarizes HF-detection results, and Fig. 5 overlays inputs, labels, and predictions. We adopt AP@IoU = 0.5 as the primary metric; it balances recall and precision, offers clinically acceptable localization tolerance for small lesions, and aligns with public benchmarks.

Table 3.

Object detection model results.

Model	AP (IoU 50%)
SSD41	0.1823
Faster-RCNN42	0.5721

Fig. 5.

The comparison of object detection algorithms is shown by identifying hyperreflective points on OCT images.

Faster-RCNN reached an AP of 0.5721, nearly triple that of SSD (0.1823). In retinal OCT, SSD’s recall on micro-HF detection was notably low, whereas Faster-RCNN maintained robust localization ability. The gap stems from their paradigms: SSD, a one-stage detector, performs dense prediction on multi-scale feature maps. While fast, its high-level features have large receptive fields and coarse resolution, and extreme foreground-background imbalance causes many false negatives and positives on pixel-sparse lesions. Faster-RCNN, a two-stage approach, first generates proposals on shallow, high-resolution features via RPN, then refines each RoI with RoI Align and a second classification-regression head, markedly reducing localization error for minute targets. Thus, under small-sample, small-pixel conditions, Faster-RCNN delivers higher recall and more stable localization.

We acknowledge that there is no unified or authoritative imaging criterion for assessing HF in patients with RVO. The “gold standard” annotations in this data set were generated through a rigorous consensus among several senior ophthalmologists, who carefully labeled each case based on raw medical images. Given that HF lesions are tiny and exhibit indistinct boundaries, their visual identification is extremely challenging. This difficulty not only complicates clinical diagnosis but also imposes stringent demands on the accuracy of AI detection models. We therefore anticipate the emergence of more advanced object detection frameworks capable of precisely localizing these minor lesions and enabling a quantitative evaluation of the state of the disease. To facilitate reproducibility and further research, the dataset publicly releases all original LabelMe JSON annotation files for use by clinicians and researchers. Experiments confirm that baseline models converge reliably on this dataset, attesting to its learnability and reliability, and underscoring its value for precise RVO assessment. By integrating HF detection and structural segmentation, the dataset enables clinical evaluation driven by artificial intelligence and focused on the macula.

However, we are aware that this study has certain limitations. First, all images were acquired on the Cirrus-HD-5000; performance on other OCT platforms may differ. Second, all patients in this study were of Asian descent. Therefore, collecting RVO lesions from more multi-center, multi-ethnic, and multi-device sources will be a trend in the future. Currently, there is no macular OCT dataset dedicated to RVO that provides both key structures and lesions; our dataset fills this gap and serves as a starting point for the quantification of fine-grained lesions.

Open-access datasets

In recent years, several OCT datasets have been published and have played an important role in advancing both clinical research and the development of AI-based analysis methods. These datasets provide valuable imaging resources and annotation standards for tasks such as retinal disease recognition, fluid identification, and structural analysis, thereby facilitating methodological progress and enabling fair benchmarking across studies.

Among them, some datasets are designed around specific retinal diseases. For example, the AROI dataset⁴³ primarily targets neovascular age-related macular degeneration and provides detailed annotations of major fluid components, including intraretinal fluid, subretinal fluid, and pigment epithelial detachment. Other datasets, such as OCT5k⁴⁴, encompass multiple diseases, including AMD and diabetic macular edema, and offer annotations for retinal layer segmentation and lesion detection. In contrast, datasets like OCTDL⁴⁵ focus mainly on image-level disease classification and are therefore more suitable for large-scale diagnostic modeling. Collectively, these datasets have established widely used benchmarks and have supported substantial progress in OCT-based analysis.

However, most existing public datasets are not specifically designed for macular edema secondary to retinal vein occlusion. Owing to the unique hemodynamic disturbances and vascular leakage patterns associated with RVO, the OCT manifestations of RVO-ME can differ from those observed in AMD or DME. Consequently, datasets constructed primarily around AMD or DME may not fully capture the structural characteristics of RVO-ME, which can limit their applicability when developing or evaluating disease-specific models.

To address this gap, the dataset constructed in this study focuses explicitly on RVO-ME and provides relatively fine-grained annotations tailored to its clinical features. These include pixel-level segmentation of intraretinal and subretinal fluid, as well as delineation of key retinal layers such as the ELM and EZ. In addition, hyperreflective foci are annotated at the region level to facilitate further pathological analysis. All images were acquired using Zeiss OCT systems, which complement existing datasets derived from other devices and contribute to cross-device data supplementation. While the dataset remains limited in scale and is based on a single-center cohort, strict annotation consistency protocols were applied to reduce subjectivity and enhance reliability. Overall, these design choices allow the dataset to more accurately reflect the structural characteristics and clinical assessment needs of RVO-ME, while also providing a focused benchmark for future methodological development.

Looking ahead, further work could evaluate model generalization and robustness by establishing a unified cross-disease and cross-device evaluation framework, for example, by training models on datasets such as AROI or OCT5k and testing them on the RVO-ME dataset. Such efforts would help to clarify model stability under varying imaging conditions and disease phenotypes, and would be an important step toward broader clinical translation, where additional challenges related to regulatory approval and workflow integration still need to be addressed.

Author contributions

F.X., G.L., and W.G. were responsible for the conceptualization and design of the study. Y.H., W.L., G.L., and F.X. were responsible for data annotation. Y.Z., Y.G.and X.X. contributed to data collection and assisted in manuscript writing. Y.H., W.G., and L.M. participated in the technical validation. W.L.and G.L. collectively supervised this research’s progress. All authors contributed to the article and approved the submitted version.

Data availability

The complete dataset is available for download via the following link: 10.6084/m9.figshare.29804435.v1.

Code availability

The code mentioned in this study can be found at https://github.com/AI-thpremed/Dual-task-baseline-RVO-ME.

Competing interests

The authors declare no competing interests.