Kvasir-SEG: A Segmented Polyp Dataset (1911.07069v1)

Abstract: 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.

• The paper presents a high-quality, open-access GI polyp segmentation dataset with pixel-wise annotations verified by medical experts.
• It compares a traditional Fuzzy C-Mean approach with a deep learning ResUNet, highlighting ResUNet's superior performance (Dice coefficient of 0.787763).
• The dataset provides a robust benchmark for developing clinical decision support systems to reduce polyp miss rates and improve early colorectal cancer detection.

Kvasir-SEG: A Segmented Polyp Dataset

The paper, "Kvasir-SEG: A Segmented Polyp Dataset," by Debesh Jha et al., presents a meticulously curated open-access dataset specifically designed for gastrointestinal (GI) polyp segmentation tasks. The dataset, annotated by medical professionals, aims to address the burgeoning need for high-quality, labeled medical images in the research sector and facilitates benchmarking and comparative evaluation of existing and new segmentation methodologies.

Introduction and Motivation

Colorectal cancer ranks alarmingly high among cancer types globally, with early detection playing a pivotal role in patient survival. Polyps, often precursors to colorectal cancer, necessitate swift identification and excision, primarily facilitated by colonoscopy. However, traditional colonoscopy is plagued with significant polyp miss rates, ranging from 14% to 30%, underscoring the imperative need for automated systems to enhance detection rates.

Dataset Construction and Verification

The Kvasir-SEG dataset builds upon the existing Kvasir collection by providing pixel-wise segmentation masks for each polyp image in a consistent high-quality format. Each image and its corresponding mask were annotated manually by a medical doctor and subsequently verified by an experienced gastroenterologist. This rigorous validation ensures the reliability and robustness of the dataset, making it highly suitable for developing and training ML and deep learning (DL) models.

Technical Contributions

The primary contributions of Kvasir-SEG are threefold:

1. Extended Dataset: The dataset significantly enhances the original Kvasir collection by including segmentation masks, bounding boxes, and corresponding metadata, providing a rich resource for multimedia and computer vision researchers.
2. Baseline Model: The authors demonstrate the utility of the dataset through a comparative paper employing a traditional Fuzzy C-Mean (FCM) clustering algorithm and a deep learning-based Residual U-Net (ResUNet) architecture. The ResUNet model, optimized for pixel-wise semantic segmentation, achieved remarkable performance on the dataset.
3. Open Access: A major highlight is the dataset's open-access nature, encouraging reproducibility and consistent benchmarking in the research community. The dataset is publicly available and can be obtained from the SimulaMet repository.

Experimental Results

The paper meticulously documents the experiments conducted with both the FCM and ResUNet models. The FCM method, which relies heavily on color as a distinguishing feature, achieved a Dice coefficient of 0.239002 and a mean Intersection over Union (IoU) of 0.314187. Conversely, the ResUNet model demonstrated a superior Dice coefficient of 0.787763 and a mean IoU of 0.777771, underscoring the effectiveness of deep learning approaches for this task.

Implications and Future Directions

The introduction of the Kvasir-SEG dataset holds substantial implications for both theoretical and practical advancements in medical image analysis. The dataset provides a critical benchmark for testing segmentation algorithms, fostering innovation and improvements in automated detection systems. Given the dataset's quality and detailed annotations, it is poised to become a cornerstone resource in developing robust clinical decision support systems aimed at reducing polyp miss rates during colonoscopies.

Conclusion

"Kvasir-SEG: A Segmented Polyp Dataset" is a significant contribution to the medical image research community, providing a well-annotated, reliable dataset that facilitates the training and evaluation of advanced computer vision models. The results obtained using the ResUNet architecture highlight the dataset's potential in enhancing state-of-the-art polyp segmentation methods, promising significant strides in clinical practice and patient outcomes. Future explorations could harness more diverse and larger datasets, aiming towards deploying these automated systems in real-world clinical scenarios.

The Kvasir-SEG dataset not only bridges a critical gap in the availability of annotated medical images but also sets a foundation for subsequent research endeavors aimed at improving diagnostic accuracy and efficiency in GI tract examinations.