task_categories:
- image-to-image
tags:
- medical
- video-super-resolution
- medical-imaging
- surgical-video
MedVSR Dataset
MedVSR is a framework and dataset collection tailored for medical video super-resolution (VSR). It addresses unique challenges in clinical medical videos, such as camera shake, noise, and abrupt frame transitions, which often lead to alignment errors in standard VSR models.
This repository contains preprocessed versions of the following medical datasets used in the study:
- HyperKvasir: Endoscopy images and videos.
- LDPolyp: A large-scale dataset for polyp detection and classification.
- EndoVis18: Data from the Endoscopic Vision Challenge.
These datasets cover various medical scenarios, including endoscopy and cataract surgeries, and are used to evaluate the MedVSR model's ability to selectively propagate consistent features and enhance tissue structures.
Dataset Preparation
According to the official repository, users should download the data from this Hugging Face repository and update the paths in the configuration files (e.g., options/medvsr_train.yml) to point to the extracted folders.
Citation
If you find this dataset or the MedVSR framework useful for your research, please cite the following paper:
@inproceedings{liu2025medvsr,
title = {MedVSR: Medical Video Super-Resolution with Cross State-Space Propagation},
author = {Liu, Xinyu and Sun, Guolei and Wang, Cheng and Yuan, Yixuan and Konukoglu, Ender},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
year = {2025}
}