Source-linked AI summary
Kvasir-SEG: A Segmented Polyp Dataset
Debesh Jha, Pia H. Smedsrud, Michael A. Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, Håvard D. Johansen
TL;DR
Annotated medical images with corresponding segmentation masks are difficult to obtain, limiting reproducible research on polyp segmentation. This paper introduces the open-access Kvasir-SEG dataset with masks and bounding boxes, and demonstrates FCM and ResUNet segmentation, with ResUNet outperforming FCM.
Problem
Annotated medical images with corresponding segmentation masks are difficult to obtain, while detecting polyps early is important because polyps can precede colorectal cancer.
Method
The paper extends Kvasir with polyp images, manually created segmentation masks and bounding boxes, and verified the annotations through gastroenterologist review before evaluating FCM and ResUNet.
Results
ResUNet achieved a Dice coefficient of 0.787763 and mean IoU of 0.777771 on the test dataset, outperforming FCM, which achieved Dice 0.239002 and mean IoU 0.314187.
Takeaways & Limitations
The open-source Kvasir-SEG dataset is intended to support reproducible evaluation and comparison of current and future computer vision methods for polyp segmentation.
Takeaways & Limitations
Some original images contain an endoscope position marking probe from the ScopeGuide.
Abstract
from arXiv · showhide
Pixel-wise image segmentation is a highly demanding task in medical-image analysis. In practice, it is difficult to find annotated medical images with corresponding segmentation masks. In this paper, we present Kvasir-SEG: an open-access dataset of gastrointestinal polyp images and corresponding segmentation masks, manually annotated by a medical doctor and then verified by an experienced gastroenterologist. Moreover, we also generated the bounding boxes of the polyp regions with the help of segmentation masks. We demonstrate the use of our dataset with a traditional segmentation approach and a modern deep-learning based Convolutional Neural Network (CNN) approach. The dataset will be of value for researchers to reproduce results and compare methods. By adding segmentation masks to the Kvasir dataset, which only provide frame-wise annotations, we enable multimedia and computer vision researchers to contribute in the field of polyp segmentation and automatic analysis of colonoscopy images.
1 Introduction
The paper motivates automated polyp segmentation by the clinical importance of early detection and the limitations of manual annotation and existing datasets. It introduces Kvasir-SEG, an open-access dataset with segmentation masks and bounding boxes, together with baseline segmentation methods.
- Motivation: Colorectal polyps can precede cancer, are frequently found during screening, and may be missed during colonoscopy.Reported polyp miss rates range from 14 to 30% depending on polyp type and size.
- Motivation: Automated pixel-wise segmentation could support faster, more accurate, and cost-effective computer-assisted interventions.Segmentation assigns labels to individual pixels, identifying regions such as polyp tissue.
- Research gap: Manual medical-image segmentation is tedious, time-consuming, expensive, and vulnerable to physician bias and inter-observer variation.Developing automated methods is also constrained by the difficulty of obtaining large, high-quality annotated medical datasets.
- Contribution: The paper provides Kvasir-SEG, extending Kvasir with polyp images, segmentation masks, and bounding boxes generated from physician annotations verified by an experienced gastroenterologist.The dataset is publicly available and open access.
- Contribution: The authors evaluate Fuzzy C-mean clustering and Deep Residual U-Net as baseline approaches for pixel-wise semantic segmentation.Dice coefficient and mean IoU are used for evaluation and proposed for future comparisons on the dataset.
2 Related Work
Prior polyp datasets provide segmentation annotations but are limited by access restrictions, small annotated subsets, or image similarity. The paper identifies a missing open-access dataset for comparable evaluation and positions Kvasir-SEG as a response.
- Available datasets: Existing polyp datasets with ground-truth segmentation masks were few, including CVC-ColonDB, ASU-Mayo, ETIS-Larib, and CVC-Clinic DB.These datasets formed the available basis for prior polyp-segmentation research.
- Available datasets: CVC-ColonDB contains 1,200 extracted frames from 15 short colonoscopy sequences, but only 300 frames are annotated and access requires registration.The annotated frames were selected to maximize visual differences.
- Available datasets: The ASU-Mayo dataset provides more than 3,500 polyp frames with masks, but its visually similar images raise an overfitting concern.Its training data comes from 20 short videos, split between videos with and without polyps.
- Available datasets: ETIS-Larib provides 196 polyp frames with masks and requires registration, whereas CVC-Clinic provides 612 masked frames from 29 colonoscopy sequences as an open-access resource.Both datasets were used in the MICCAI 2015 automatic polyp-detection sub-challenge.
- Research gap: The literature review concludes that an open-access dataset for comparable evaluation was missing, motivating the extension of Kvasir with segmentation masks.The authors present the resulting data and baseline work as resources for developing robust and efficient systems.
3 The Kvasir-SEG dataset
Kvasir-SEG extends the Kvasir gastrointestinal image dataset with 1,000 polyp images, pixel-wise masks, and bounding boxes for segmentation research. An engineer and medical doctor manually annotated the images, with review by an experienced gastroenterologist.
- Kvasir-SEG builds on Kvasir, a multiclass gastrointestinal dataset whose original collection contains 8,000 images across eight classes.
- The dataset addresses Kvasir’s frame-wise annotation limitation by adding annotated polyp images with corresponding segmentation masks.
- Polyp regions are represented as white foreground pixels and non-polyp background as black pixels in the masks.
- Some original images include an endoscope position-marking probe from the Olympus ScopeGuide.
- Each dataset folder contains 1,000 JPEG images or masks, while bounding-box coordinates are stored in a JSON file using matching filenames.
- A team consisting of an engineer and medical doctor manually outlined all polyp margins, and an experienced gastroenterologist reviewed the annotations.
4 Suggested Metrics
The paper uses Dice coefficient and Intersection over Union to evaluate pixel-wise agreement between predicted polyp segmentations and ground truth. IoU is additionally evaluated across threshold values.
- Dice coefficient and mean IoU are used to evaluate pixel-wise polyp-versus-non-polyp segmentation performance.
- Dice coefficient: Dice coefficient compares predicted segmentation pixels with the ground-truth object pixels.
- Dice coefficient: Dice coefficient is computed from true positives, false positives, and false negatives using the predicted and ground-truth pixel sets.
- Intersection over Union: IoU measures similarity between the predicted pixel set and its corresponding ground-truth pixel set.
- Intersection over Union: At each threshold t, the IoU-related calculation produces a precision value based on predicted objects and ground-truth objects.
5 Evaluation
The evaluation compares an unsupervised FCM clustering baseline with a ResUNet deep-learning baseline on Kvasir-SEG. ResUNet achieves substantially stronger test-set segmentation metrics, while qualitative and learning-curve results further characterize performance.
- Evaluation setup: The experiments assess Kvasir-SEG using efficient unsupervised FCM clustering and a residual-block ResUNet architecture.The evaluation uses two baseline methods to demonstrate the dataset’s utility for polyp segmentation.
- FCM baseline: The FCM pipeline preprocesses images, reshapes them into one-dimensional inputs, and reshapes its output into two-dimensional binary masks.Preprocessing includes grayscale conversion, median blurring, thresholding, edge extraction, and dilation before clustering.
- ResUNet baseline: ResUNet training uses augmentation, 320 × 320 images, an 80%/10%/10% train-validation-test split, and five convolutional blocks in both encoder and decoder.Augmentation includes flipping, random crops, scaling, rotation, brightness changes, cutout, and random erasing.
- Quantitative results: 0.787763 Dice coefficient and 0.777771 mean IoU were achieved by ResUNet on the test dataset, compared with 0.239002 Dice coefficient and 0.314187 mean IoU for FCM.The reported ResUNet scores include training, validation, and testing results in Table 1.
- Qualitative results: Qualitative comparisons show original images, ground truth masks, FCM outputs, and ResUNet outputs in separate figure columns.Figure 3 is used alongside Table 1 to compare the two segmentation approaches.
- Learning and scope: ResUNet’s learning curve reports Dice coefficient against training epochs, with the best model selected for testing on previously unseen data.The paper describes the presented results as good but calls for more research before clinic-applicable performance is achieved.
- Discussion: The study attributes FCM’s weaker performance partly to reliance on color, since polyps and other gastrointestinal conditions can have similar appearances.The paper also notes that FCM does not use data augmentation or learned parameters, whereas ResUNet does.
6 Conclusion
The paper presents Kvasir-SEG as an open-access polyp segmentation dataset and demonstrates its use with FCM clustering and ResUNet. The dataset is intended to support reproducible evaluation and comparison of computer-vision methods.
- Kvasir-SEG is a new polyp segmentation dataset developed to support extensive and reproducible research.
- The paper presents both an FCM clustering algorithm and a ResUNet-based approach for automatic polyp segmentation.
- The results show that ResUNet outperforms FCM clustering.
- Kvasir-SEG is released as open-source to help researchers evaluate and compare existing and future computer-vision methods.The authors connect this aim to progress toward clinically acceptable computer-assisted intervention methods.