All Open Datasets
The Kvasir-SEG Dataset
Pixel-wise image segmentation is a highly demanding task in medical image analysis. It is difficult to find annotated medical images with corresponding segmentation mask. Here, we present Kvasir-SEG. It is an open-access dataset of gastrointestinal polyp images and corresponding segmentation masks, manually annotated and verified by an experienced gastroenterologist. This work will be valuable for researchers to reproduce results and compare their methods in the future. By adding segmentation masks to the Kvasir dataset, which until today only consisted of framewise annotations, we enable multimedia and computer vision researchers to contribute in the field of polyp segmentation and automatic analysis of colonoscopy videos.


The human gastrointestinal (GI) tract is made up of different sections, one of them being the large bowel. Several types of anomalies and diseases can affect the large bowel, such as colorectal cancer. Colorectal cancer is the second most common cancer type among women and third most common among men. Polyps are precursors to colorectal cancer, and is found in nearly half of the individuals at age 50 having a screening colonoscopy, and are increasing with age. Colonoscopy is the gold standard for detection and assessment of these polyps with subsequent biopsy and removal of the polyps. Early disease detection has a huge impact on survival from colorectal cancer, and polyp detection is therefore important. In addition, several studies have shown that polyps are often overlooked during colonoscopies, with polyp miss rates of 14%-30% depending on the type and size of the polyps. Increasing the detection of polyps has been shown to decrease risk of colorectal cancer. Thus, automatic detection of more polyps at an early stage can play a crucial role in improving both prevention of and survival from colorectal cancer. This is the main motivation behind the development of a Kvasir-SEG dataset.

Original Kvasir Dataset Details

The Kvasir dataset comprises 8000 gastrointestinal (GI) tract images, each class consisting of 1000 images. These images were collected and verified by experienced gastroenterologist from Vestre Viken Health Trust in Norway. The eight classes of the dataset include anatomical landmarks, pathological findings and endoscopic procedures. Each class' images are saved in a separate folder corresponding to the class they belong to. More detailed explanation about each image class, data collection procedure and dataset details can be found on the Kvasir dataset's homepage at

Kvasir-SEG Dataset Details

The Kvasir-SEG dataset (size 46.2 MB) contains 1000 polyp images and their corresponding ground truth from the Kvasir Dataset v2. The resolution of the images contained in Kvasir-SEG varies from 332x487 to 1920x1072 pixels. The images and its corresponding masks are stored in two separate folders with the same filename. The image files are encoded using JPEG compression, and online browsing is facilitated. The open-access dataset can be easily downloaded for research and educational purposes.

The bounding box (coordinate points) for the corresponding images are stored in a JSON file. This dataset is designed to push the state of the art solution for the polyp detection task.
Some examples of the dataset.

Image Ground Truth Bounding Box

Applications of the Dataset

The Kvasir-SEG dataset is intended to be used for researching and developing new and improved methods for segmentation, detection, localization, and classification of polyps. Multiple datasets are prerequisites for comparing computer vision-based algorithms, and this dataset is useful both as a training dataset or as a validation dataset. These datasets can assist the development of state-of-the-art solutions for images captured by colonoscopes from different manufacturers. Further research in this field has the potential to help reduce the polyp miss rate and thus improve examination quality. The Kvasir-SEG dataset is also suitable for general segmentation and bounding box detection research. In this context, the datasets can accompany several other datasets from a wide range of fields, both medical and otherwise.

Ground Truth Extraction

We uploaded the entire Kvasir polyp class to Labelbox and created all the segmentations using this application. The Labelbox is a tool used for labeling the region of interest (ROI) in image frames, i.e., the polyp regions for our case. We manually annotated and labeled all of the 1000 images with the help of medical experts. After annotation, we exported the files to generate masks for each annotation. The exported JSON file contained all the information about the image and the coordinate points for generating the mask. To create a mask, we used ROI coordinates to draw contours on an empty black image and fill the contours with white color. The generated masks are a 1-bit color depth images. The pixels depicting polyp tissue, the region of interest, are represented by the foreground (white mask), while the background (in black) does not contain positive pixels. Some of the original images contain the image of the endoscope position marking probe, ScopeGuide TM, Olympus Tokyo Japan, located in one of the bottom corners, seen as a small green box. As this information is superfluous for the segmentation task, we have replaced these with black boxes in the Kvasir-SEG dataset.

Suggested Metrics

There are different metrics for evaluating the performance of the architectures on the image segmentation dataset. For medical image segmentation task, the most commonly used ones are Dice coefficient and Intersection over Union (IOU). Based on related work in this field, we have used these metrics for the evaluation of the algorithms. In future work, we encourage the use of these metrics for evaluating the performance of the model. In the future, it might be even better to include as many as possible metrics for the fair comparison of the models.

Dice coefficient is a standard metric for comparing the pixel-wise results between predicted segmentation and ground truth. It is defined as:

The Intersection over Union (IoU) is another standard metric to evaluate a segmentation method. The IoU calculates the similarity between predicted (A) and its corresponding ground truth (B) as shown in the equation below:
In the equations above:

  • A: Predicted set of pixels
  • B: Ground truth of the object
  • TP: True Positive
  • FP: False Positive
  • FN: False Negative

Use terms and citation

The use of the Kvasir-SEG dataset is restricted for research and educational purposes. The use of the Kvasir-SEG dataset for commercial purposes is forbidden without prior written permission. For other purposes, contact us (see below). In all documents and publications that use the Kvasir-SEG dataset or report experimental results based on the Kvasir-SEG dataset, a reference to the dataset paper has to be included (see below). Please email if you have any questions regarding how to cite the dataset.

Debesh Jha, Pia H. Smedsrud, Michael A. Riegler, Pål Halvorsen, Dag Johansen, Thomas de Lange, and Håvard D. Johansen, Kvasir-SEG: A Segmented Polyp Dataset, In Proceedings of the ternational conference on Multimedia Modeling, Republic of Korea, 2020.



title={Kvasir-seg: A segmented polyp dataset},
author={Jha, Debesh and Smedsrud, Pia H and Riegler, Michael A and Halvorsen, P{\aa}l and
de Lange, Thomas and Johansen, Dag and Johansen, H{\aa}vard D},
booktitle={International Conference on Multimedia Modeling},


Kvasir-SEG Dataset The latest version of Kvasir-SEG is downloadable here. For details, see our paper.


The Kvasir-SEG dataset includes 196 polyps smaller than 10 mm classified as Paris class 1 sessile or Paris class IIa. We have selected it with the help of expert gastroenterologists. We have released this dataset separately as a subset of Kvasir-SEG. We call this subset Kvasir-Sessile.

The dataset is publicly available. It can be downloaded from here:

If you use this dataset, please cite our paper,


title={A Comprehensive Study on Colorectal Polyp Segmentation with ResUNet++,
Conditional Random Field and Test-Time Augmentation},
author={Jha, Debesh and Smedsrud, Pia H and Johansen, Dag and de Lange, Thomas and
Johansen, H{\aa}vard D and Halvorsen, P{\aa}l and Riegler, Michael A},


Email debesh (_at_) simula (_dot_) no if you have any questions about the dataset and our research activities. We always welcome collaboration and joint research!