Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Data volume 13, Article number: 528 (2026)
2631
Metrics details
Cystoscopy is a common endoscopic procedure used for the visual inspection of the lower urinary tract, particularly for the detection and surveillance of bladder cancer. Artificial intelligence (AI) strategies could help address the recognized shortcomings of cystoscopy by identifying malignant and non-malignant regions of interest (ROIs) and providing real-time clinical decision support for biopsy and surgical resection. However, curating a dataset for training AI models is challenging and time-time consuming. We present CystoDS, a high-quality bladder imaging dataset derived from standard white light cystoscopy that is ready for AI applications for detection of bladder cancer and cancer-mimicking benign lesions. This dataset includes 8,067 images from 160 patients labelled with five classes and 22 subclasses, along with segmentation data for 768 of the images. We detail the methods used for image acquisition, the structure of the dataset, and our technical validation process to demonstrate the AI-readiness of the data.
Bladder cancer is the 6th most commonly diagnosed malignancy in the United States and disproportionately affects males, with 63,070 new cases in males compared to 20,120 new cases in females in 20241. White light cystoscopy is the standard endoscopic method for visual identification of suspected bladder cancer and other regions of interest (ROI) in the bladder2. In a clinic setting, a flexible cystoscope is inserted through the urethra into the bladder to visualize the mucosa and identify any abnormalities. If tumors or ROIs are detected, an endoscopic surgical procedure called transurethral resection of bladder tumor (TURBT) is performed in the operating room to treat the tumor and establish a pathologic diagnosis3.
Standard white light cystoscopy has several recognized shortcomings including difficulty in differentiating cancer from non-cancerous ROIs,4,5 detecting non-papillary carcinoma in situ (CIS) that lacks visually distinct borders6,7,8, enumerating multifocal tumors, and ensuring complete tumor resection9. Bladder cancer has the highest lifetime treatment costs of any cancer10, highlighting the need to address these shortcomings. Efforts in this space include enhanced cystoscopy technologies such as blue light cystoscopy, which requires instillation of the optical imaging agent hexaminolevulinate11. Currently, blue light cystoscopy is recommended by recognized guidelines, if available, to increase tumor detection12.
Artificial intelligence (AI) and deep learning have the potential to enhance visualization, enabling surgeons to better identify, evaluate, and treat bladder cancer. We previously reported CystoNet, an AI model that accurately detects bladder tumors during cystoscopy13,14. While this model demonstrated promising performance for papillary tumors, to create more comprehensive and accurate AI tools, a large dataset of quality images following FAIR principles (Findable, Accessible, Interoperable, Reusable) is required for model training and testing15. As new deep learning techniques become available, having AI-ready datasets with accurate, complete, and consistent data becomes imperative for model training and validation16. While several large endoscopic imaging datasets exist for other organ systems, such as the gastrointestinal tract17,18,19, large cystoscopy imaging datasets for training AI models are lacking. One bladder endoscopy dataset used to train a classification model has been published20. Another small cystoscopy imaging atlas derived from a textbook was published, but it is limited due the small size of the dataset and suboptimal image quality21.
In this study, we present CystoDS22, an accessible cystoscopy image dataset containing 8,067 labelled images from 160 patients. These labelled images include pathologically confirmed malignant and non-malignant ROIs, anatomical landmarks of the bladder, foreign bodies, and normal bladder mucosa. We performed technical validation using our labelled dataset to demonstrate the feasibility and potential usability of CystoDS for AI model training and validation. The benchmark performance of CystoDS was validated using publicly available deep learning image classification models towards developing AI-driven solutions to enhance endoscopic bladder tumor detection and resection.
With Institutional Review Board (IRB #29838 and #36085) approval, human subjects (n = 160) undergoing cystoscopy and TURBT were consented and enrolled from 2016 to 2022 at the Veterans Affairs Palo Alto Health Care System (VAPAHCS). Patients consented to recording of standard of care cystoscopies to be uploaded to a secure database as de-identified files that can be shared electronically. Informed consent was obtained at the time of cystoscopy by trained research personnel. Only the protocol director and his trained research personnel have full access to patient data. Table 1 shows patient demographics and diversity of images and ROIs. Video and image data were acquired at the time of TURBT (n = 200, KARL STORZ Endoskope, Tuttlingen, Germany) in the operating room or flexible cystoscopy (n = 8, Olympus Corporation, Tokyo, Japan) in the clinic. Additional still images were captured from cystoscopy videos using built-in MacOS screenshotting software and FFmpeg, an open-source video processing software (ffmpeg.org). A schematic of data acquisition and storage is shown in Fig. 1.
CystoDS imaging data acquisition and storage. A total of 8,067 images were captured from cystoscopy videos. Images were then labelled with relevant clinical and pathological information retrieved from medical records. 768 images containing regions of interest were additionally segmented and reviewed by expert urologists and then stored with corresponding segmentation data in JSON format.
All images (n = 8,067) were labelled after clinical and pathological diagnoses were established. The labelling process involves renaming image file names with clinical information following a standardized protocol23. This clinical information served as the basis of the class and subclass determination for each image. File names of images included in CystoDS were changed to a random 8-digit alphanumeric sequence to protect patient privacy and only relevant de-identified clinical data were stored in the accompanying cystods.csv file.
ROIs in select images (n = 768, 9.6% of total images) were segmented by an expert urologist using LabelMe. Segmentation consisted of outlining ROIs considered suspicious enough to warrant a biopsy and/or surgical resection for pathological diagnosis. Segmented images were then reviewed by another expert urologist for quality control. Data of verified segmentations were stored in JSON files.
The complete CystoDS dataset22, including images, segmentation data, and metadata, is available from Open Science Framework (OSF) at https://osf.io/xvdhy/. CystoDS comprises three main components: the “image” folder, “segmentation” folder, and “cystods.csv” file. The “image” folder contains all images included in the dataset (n = 8,067) in PNG format, with resolutions ranging from 252 to 5120 pixels in width and 209 to 2880 pixels in height. The median file size is 95 kilobytes (KB), with a range of 44 to 8787 KB. The “segmentations” folder contains JSON files that store segmentation data for the segmented images (n = 768). Similar to other published endoscopy datasets, not all images in CystoDS have associated segmentation data (Table 2); however, all images in CystoDS have a class label. CystoDS has a comparable number of segmented and labelled images to existing endoscopy datasets such as HyperKvasir17,24 (n = 10,662 labelled images; 1,000 segmented images), PolypGen18,25 (n = 8,037 labelled images; 1,537 segmented images), and GastroVision19,26 (n = 8,000 labelled images; 0 segmented images). Compared to the bladder endoscopy dataset published by Lazo et al.20,27 (n = 1,754 labelled images; 0 segmented images; 23 patients), CystoDS offers images from a greater number of patients (n = 160) and segmentation data for a subset of the dataset. Additionally, CystoDS contains a higher percentage (79%) of normal mucosa images. This skew towards normal mucosa images reflects the reality of bladder endoscopy where typically more normal mucosa is visualized compared to abnormal. Users of our dataset can address the potential bias that comes with class imbalance by randomly selecting a subset of normal mucosa images during training.
The file, cystods.csv, contains the metadata for all 8,067 images including variables such as filename, participant identification number (pid), visit, ROI, multifocal, bladder cancer status (bca), class, subclass, subclass2, stage, morphology, modality, and JSON. The “filename” variable is a randomized image file name that was generated using the ids package in R to deidentify any protected health information (PHI) in the original file name. Every participant was assigned a random patient identification number (pid) after obtaining informed consent and prior to data collection. Due to the recurrent nature of bladder cancer, participants often undergo multiple cystoscopies for disease surveillance and therefore contributed to the dataset multiple times. Specifically, there are images from one (n = 83 patients) or two (n = 57) visits, with some contributing images from three (n = 14), four (n = 3), five (n = 2), and seven (n = 1) visits. Visits were numbered sequentially for each individual and stored in the “visit” variable. Within any given visit, multiple ROIs may be identified, and multiple images of the same ROI from different perspective can be generated. All images are assigned a “class” and “subclass”: the five classes – malignant, non-malignant, anatomical landmarks, foreign bodies, and normal mucosa are further defined into 22 subclasses (Figs. 2, 3). The “ROI” variable denotes the identification number for the given image, with different numbers represent different ROIs or identify the same ROI from different perspectives. In cases where enumeration of ROIs was not feasible, or multiple ROIs were sent for pathology as one specimen, images were labelled “multifocal” instead of a number for their ROI variable (Fig. 4b). Additionally, the “multifocal” variable is assigned “2–7” or “8+” as an estimate of the number of tumors associated with the multifocality. The “bca” variable is set to 1 for images containing bladder cancer and 0 when no cancer is present in the image.
Malignant subclasses (n = 4): LowGradePapillary, HighGradePapillary, CIS, and PreMalignant are subclasses based on guidelines for non-muscle invasive bladder cancer (NMIBC)28. PreMalignant includes urothelial tumors that are not assigned a grade or stage such as urothelial proliferation of undetermined malignant potential (UPUMP) and papillary urothelial neoplasms of low malignant potential (PUNLMP)29.
Non-malignant subclasses (n = 8): BenignNOS (benign, not otherwise specified), InflammationNOS (inflammation, not otherwise specified), CCG (cystitis cystica and glandularis), Denuded (denuded urothelium), UrothelialPapilloma, SquamousMetaplasia, NephrogenicAdenoma, and BenignRare are subclasses determined through internal discussion between expert urologists and pathologists to consolidate various benign ROIs that can mimic bladder cancer on cystoscopy.
BenignNOS includes pathological findings such as reactive changes, urothelial mucosa, atypia, and dysplasia. InflammationNOS includes the various non-specific diagnoses of cystitis, while CCG is a more specific diagnosis associated with chronic inflammation. BenignRare includes rare cystoscopic findings such as malakoplakia and melanosis.
Anatomical landmarks subclasses (n = 6): UreteralOrifice, ResectionBed, ResectionScar, ProstaticUrethra, Trabeculation, and Diverticulum are common anatomical landmarks observed in patients at risk for bladder cancer30.
Foreign bodies subclasses (n = 4): AirBubble, ResectionLoop, BiopsyForcep, and Stent are common cystoscopy findings and tools.
Normal mucosa: Images of visually normal-appearing bladder tissue are assigned this class and NA (not applicable) for subclass.
CystoDS classes, subclasses and representative images. A total of 8067 images from 160 patients were divided into five classes: (a) malignant; (b) non-malignant; (c) anatomical landmarks; (d) foreign bodies; and (e) normal mucosa. Malignant includes pathologically confirmed papillary urothelial carcinoma, carcinoma in situ (CIS), and a pre-malignant urothelial tumor. Non-malignant are cancer-mimicking in appearance but confirmed pathologically to be benign. Anatomical landmarks are structures (e.g. ureteral orifice) or features within the bladder as determined by experienced urologists. Foreign bodies include surgical instruments and other findings that are not normally present in the bladder. Normal mucosa is visually normal-appearing bladder tissue. Abbreviations: NOS (not otherwise specified), CCG (cystitis cystica and glandularis).
CystoDS images by class and subclass. The 22 subclasses are grouped into five classes: Malignant (n = 998), Non-malignant (n = 221), Anatomical landmarks (n = 211), Foreign bodies (n = 251). Normal mucosa images (n = 6,386) are not shown.
A subset of cases contains ROIs with complex pathology. Twelve ROIs are pathologically low grade with focal high grade papillary urothelial carcinoma. In these cases, the images were assigned class Malignant, subclass LowGradePapillary, and HighGradePapillary for “subclass2”. In a four cases CIS was present in addition to papillary urothelial carcinoma. These images were assigned class Malignant, subclass HighGradePapillary, and CIS for subclass2. Bladder cancer pathology also includes local stage information (T0, Ta, Tis, T1, T2) for depth of tumor involvement. These data are stored in the “stage” variable. Bladder cancer pathology is defined by both the grade and stage. The 998 images of malignant ROIs in the dataset fall into five combinations of grade and stage: LG (low grade) Ta (n = 491), LG T1 (n = 2), HG (high grade) Ta (n = 210), HG T1 (n = 170), and HG T2 (n = 53). Images that do not contain malignant ROIs with stage information are assigned NA for stage. “Morphology” refers to the physical appearance of an ROI. Malignant and non-malignant ROIs are assigned either papillary or non-papillary morphology (Fig. 4). In some cases, blue light cystoscopy was used in addition to white light cystoscopy. Images are assigned either BLC (n = 450) or WLC (n = 7,617) respectively under “modality”. If an image has an associated segmentation file, “JSON” is set to 1; otherwise, it is set to 0 (Fig. 5).
Representative images of bladder tumor morphology and multifocality. Morphology and multifocality are additional labels included in the metadata for the dataset. (a) ROIs that appear papillary (n = 875) or non-papillary (n = 344) can be either malignant or non-malignant. The malignant LowGradePapillary ROI depicted in a.1 appear morphologically similar to the non-malignant UrothelialPapilloma shown in a.3. Similarly, the CIS in a.2 and BenignNOS are difficult to differentiate. (b) Images labelled as multifocal (n = 100) contain multiple malignant tumors within a single image. This creates a challenge in enumerating tumors within a bladder.
768 segmented images in CystoDS by class and subclass. Segmented images are images with a corresponding segmentation mask that was created by a urologist and reviewed by a second urologist. Normal mucosa images were not segmented.
We performed technical validation of the CystoDS dataset using publicly available deep learning models to demonstrate its reliability and suitability for bladder ROI classification tasks.
We trained four widely used deep learning models for medical image classification to establish benchmark performance on the dataset. These models were selected to evaluate both standard convolutional neural network (CNN) architectures and advanced transformer-based designs, providing a comprehensive basis for comparison. ResNet31 is a CNN known for its straightforward design and robust feature extraction capabilities. ResNeXt32 extends ResNet with a grouped convolution strategy, balancing efficiency and complexity to handle diverse ROI types. HRNet33 preserves high-resolution representations throughout the network, making it particularly effective for capturing fine-grained details and spatial accuracy in medical imaging tasks. Swin-Transformer34, a state-of-the-art vision transformer, utilizes a hierarchical self-attention mechanism with shifted windows, offering superior performance by modelling both local and global features.
To enable training, validation, and internal testing for classification of regions of interest (ROIs), we randomized the CystoDS dataset into three subsets following a 70:15:15 split. Randomization was performed at the patient level ensuring that no patient’s data appeared in more than one subset. The final data splits include a training set of 1,772 images from 128 patients for model training, a validation set of 226 images from 15 patients for hyperparameter tuning and early stopping, and a test set of 219 images from 17 patients for final model evaluation. Furthermore, we simplified the classification task for our model by grouping image classes into two categories: ROI and non-ROI. The ROI group (n = 1,215) included malignant images (n = 994) and non-malignant images (n = 221). The non-ROI group (n = 1,002) consisted of anatomical landmarks (n = 421), foreign bodies (n = 41), and a randomly selected subset of normal mucosa (n = 540) from 6,386 available images. Since normal mucosa was the largest class, limiting its inclusion to 540 images (~10%) helped balance the dataset and minimize potential model bias.
To further assess generalizability of CystoDS, we performed external validation by applying the models trained on CystoDS images to the independent bladder endoscopy dataset published by Lazo et al.20,27, without any fine-tuning.
To evaluate ROI classification performance of the models with the CystoDS dataset, standard metrics including sensitivity, specificity, accuracy, precision, and F-1 score were determined.
Table 3 shows the variation in standard metric across the different deep learning models for internal and external validation. Trained on the CystoDS dataset, Swin-Transformer achieved the highest performance across all metrics, with an accuracy of 0.831 and 0.873 in internal and external validation, respectively and F1-scores of 0.856 and 0.862 in internal and external validation, respectively. This superior performance highlights the potential of hierarchical attention mechanisms to address the diverse characteristics of CystoDS, positioning it as a promising avenue for future research. With the external validation set the models exhibited varying levels of cross-dataset performance, with some demonstrating strong generalizability.
The variability in model performance further emphasizes the importance of tailoring model selection to specific clinical objectives. Moreover, the dataset’s high-quality segmentations and detailed categorization enable researchers to evaluate model performance across a wide range of ROIs and non-ROIs, driving advancements in diagnostic accuracy and clinical utility.
The complete CystoDS dataset22, including images, segmentation data, and metadata, is available from Open Science Framework (OSF) at https://doi.org/10.17605/OSF.IO/XVDHY.
The code used to test the dataset is available on GitHub at https://github.com/liaolabsu/CystoDS.
Siegel, R. L., Giaquinto, A. N. & Jemal, A. Cancer statistics, 2024. CA A Cancer J Clinicians 74, 12–49 (2024).
Google Scholar
Ahmadi, H., Duddalwar, V. & Daneshmand, S. Diagnosis and Staging of Bladder Cancer. Hematology/Oncology Clinics of North America 35, 531–541 (2021).
Article PubMed Google Scholar
Cheung, G., Sahai, A., Billia, M., Dasgupta, P. & Khan, M. S. Recent advances in the diagnosis and treatment of bladder cancer. BMC Med 11, 13 (2013).
Article PubMed PubMed Central Google Scholar
Hameed, O. & Humphrey, P. A. Pseudoneoplastic Mimics of Prostate and Bladder Carcinomas. Archives of Pathology & Laboratory Medicine 134, 427–443 (2010).
Article Google Scholar
Samaratunga, H., Delahunt, B., Yaxley, J. & Egevad, L. Tumour-like lesions of the urinary bladder. Pathology 53, 44–55 (2021).
Article PubMed Google Scholar
Erton, M., Ilker, Y. & Akdaş, A. Carcinoma in situ and treatment options. International Urology and Nephrology 28, 33–42 (1996).
Article CAS PubMed Google Scholar
van der Meijden, A. et al. Significance of bladder biopsies in Ta,T1 bladder tumors: a report from the EORTC Genito-Urinary Tract Cancer Cooperative Group. EORTC-GU Group Superficial Bladder Committee. Eur Urol 35, 267–271 (1999).
Article PubMed Google Scholar
Herr, H. W., Al-Ahmadie, H., Dalbagni, G. & Reuter, V. E. Bladder cancer in cystoscopically normal-appearing mucosa: a case of mistaken identity? BJU Int 106, 1499–1501 (2010).
Article PubMed Google Scholar
Hermann, G. G., Mogensen, K., Carlsson, S., Marcussen, N. & Duun, S. Fluorescence-guided transurethral resection of bladder tumours reduces bladder tumour recurrence due to less residual tumour tissue in T a/T1 patients: a randomized two-centre study. BJU International 108, E297–E303 (2011).
Article PubMed Google Scholar
Evolution and patterns of global health financing 1995–2014. development assistance for health, and government, prepaid private, and out-of-pocket health spending in 184 countries. Lancet 389, 1981–2004 (2017).
Article Google Scholar
Shkolyar, E. et al. Optimizing cystoscopy and TURBT: enhanced imaging and artificial intelligence. Nat Rev Urol 22, 46–54 (2025).
Article PubMed Google Scholar
Holzbeierlein, J. M. et al. Diagnosis and Treatment of Non-Muscle Invasive Bladder Cancer: AUA/SUO Guideline: 2024 Amendment. J Urol. 211, 533–538 (2024).
Article PubMed Google Scholar
Shkolyar, E. et al. Augmented Bladder Tumor Detection Using Deep Learning. European Urology 76, 714–718 (2019).
Article PubMed PubMed Central Google Scholar
Chang, T. C. et al. Real-time Detection of Bladder Cancer Using Augmented Cystoscopy with Deep Learning: a Pilot Study. Journal of Endourology end.2023.0056, https://doi.org/10.1089/end.2023.0056 (2023).
Huerta, E. A. et al. FAIR for AI: An interdisciplinary and international community building perspective. Sci Data 10, 487 (2023).
Article CAS PubMed PubMed Central Google Scholar
Ng, M. Y. et al. Perceptions of Data Set Experts on Important Characteristics of Health Data Sets Ready for Machine Learning. JAMA Netw Open 6, e2345892 (2023).
Article PubMed PubMed Central Google Scholar
Borgli, H. et al. HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Sci Data 7, 283 (2020).
Article PubMed PubMed Central Google Scholar
Ali, S. et al. A multi-centre polyp detection and segmentation dataset for generalisability assessment. Sci Data 10, 75 (2023).
Article PubMed PubMed Central Google Scholar
Jha, D. et al. GastroVision: A Multi-class Endoscopy Image Dataset for Computer Aided Gastrointestinal Disease Detection. In: Maier, A.K., Schnabel, J.A., Tiwari, P., Stegle, O. (eds) Machine Learning for Multimodal Healthcare Data. ML4MHD 2023. Lecture Notes in Computer Science, vol 14315. Springer, Cham. https://doi.org/10.1007/978-3-031-47679-2_10 (2023).
Lazo, J. F. et al. Semi-Supervised Bladder Tissue Classification in Multi-Domain Endoscopic Images. IEEE Trans. Biomed. Eng. 70, 2822–2833 (2023).
Article PubMed ADS Google Scholar
Eminaga, O. et al. Efficient Augmented Intelligence Framework for Bladder Lesion Detection. JCO Clinical Cancer Informatics e2300031 https://doi.org/10.1200/CCI.23.00031 (2023).
Lee, T. J. et al. CystoDS. Open Science Framework https://doi.org/10.17605/OSF.IO/XVDHY (2025).
Eminaga, O. et al. Conceptual Framework and Documentation Standards of Cystoscopic Media Content for Artificial Intelligence. J Biomed Inform 142, 104369 (2023).
Article PubMed PubMed Central Google Scholar
Borgli, H. et al. The HyperKvasir Dataset. Open Science Framework https://doi.org/10.17605/OSF.IO/MH9SJ (2019).
Ali, S. et al. PolypGen. Synapse https://doi.org/10.7303/syn26376615.
Jha, D., Sharma, V., Riegler, M. A., Halvorsen, P. & Bagci, U. GastroVision. Open Science Framework (2023).
Lazo, J. F. et al. Endoscopic Bladder Tissue Classification Dataset. Zenodo https://doi.org/10.5281/zenodo.7741476 (2023).
Woldu, S. L., Bagrodia, A. & Lotan, Y. Guideline of Guidelines – Non-Muscle Invasive Bladder Cancer. BJU Int 119, 371–380 (2017).
Article PubMed PubMed Central Google Scholar
Netto, G. J. et al. The 2022 World Health Organization Classification of Tumors of the Urinary System and Male Genital Organs—Part B: Prostate and Urinary Tract Tumors. European Urology 82, 469–482 (2022).
Article PubMed Google Scholar
Engelsgjerd, J. S. & Deibert, C. M. Cystoscopy. in StatPearls (StatPearls Publishing, Treasure Island (FL), 2025).
He, K., Zhang, X., Ren, S. & Sun, J. Deep Residual Learning for Image Recognition. in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 770–778, https://doi.org/10.1109/CVPR.2016.90 (IEEE, Las Vegas, NV, USA, 2016).
Xie, S., Girshick, R., Dollar, P., Tu, Z. & He, K. Aggregated Residual Transformations for Deep Neural Networks. in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 5987–5995, https://doi.org/10.1109/CVPR.2017.634 (IEEE, Honolulu, HI, 2017).
Sun, K., Xiao, B., Liu, D. & Wang, J. Deep High-Resolution Representation Learning for Human Pose Estimation. in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 5686–5696, https://doi.org/10.1109/CVPR.2019.00584 (IEEE, Long Beach, CA, USA, 2019).
Liu, Z. et al. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) 9992–10002, https://doi.org/10.1109/ICCV48922.2021.00986 (IEEE, Montreal, QC, Canada, 2021).
Download references
The authors gratefully acknowledge research support from NIH R01 CA260426 (J. Liao and L. Xing) and Department of Veterans Affairs BLR&D I01 BX005598 (J. Liao).
Department of Urology, Stanford University School of Medicine, Stanford, USA
Timothy J. Lee, Liang Qiu, Kathleen E. Mach, Dylan Peterson, Qingsong Yao, Eugene Shkolyar & Joseph C. Liao
Veterans Affairs Palo Alto Health Care System, Livermore, USA
Timothy J. Lee, Liang Qiu, Kathleen E. Mach, Qingsong Yao, Eugene Shkolyar & Joseph C. Liao
Center for Artificial Intelligence in Medicine & Imaging, Stanford University School of Medicine, Stanford, USA
Jin Long, Lei Xing & Joseph C. Liao
Department of Radiation Oncology, Stanford University School of Medicine, Stanford, USA
Md Tauhidul Islam & Lei Xing
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
Study design was conceived by T.J.L., L.Q., J.L. and J.C.L. T.J.L., L.Q. and JL prepared and cleaned the data for publication. E.S. and J.C.L. procured the segmentation data. L.Q. and Q.Y. performed the technical validation. All authors provided critical review of the manuscript and agreed to submission.
Correspondence to Joseph C. Liao.
The authors declare no competing interests.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Reprints and permissions
Lee, T.J., Qiu, L., Long, J. et al. CystoDS: a multiclass endoscopy image dataset for artificial intelligence-assisted bladder cancer detection. Sci Data 13, 528 (2026). https://doi.org/10.1038/s41597-026-06887-z
Download citation
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s41597-026-06887-z
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Data (Sci Data)
ISSN 2052-4463 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing: Cancer newsletter — what matters in cancer research, free to your inbox weekly.
Leave a Reply