[Paper Review] WCEbleedGen: A wireless capsule endoscopy dataset and its benchmarking for automatic bleeding classification, detection, and segmentation

Palak Handa, Manas Dhir+5 more|arXiv (Cornell University)|Aug 22, 2024

Gastrointestinal Bleeding Diagnosis and TreatmentMedicine3 citations

TL;DR

This paper introduces WCEbleedGen, a high-quality, open-source, medically annotated dataset of 2,618 wireless capsule endoscopy frames for automatic bleeding classification, detection, and segmentation. It benchmarks nine classification, three detection, and three segmentation deep learning models, with VGG19, YOLOv8n, and LinkNet achieving the best performance, setting a new standard for multi-task AI in WCE analysis.

ABSTRACT

Computer-based analysis of Wireless Capsule Endoscopy (WCE) is crucial. However, a medically annotated WCE dataset for training and evaluation of automatic classification, detection, and segmentation of bleeding and non-bleeding frames is currently lacking. The present work focused on development of a medically annotated WCE dataset called WCEbleedGen for automatic classification, detection, and segmentation of bleeding and non-bleeding frames. It comprises 2,618 WCE bleeding and non-bleeding frames which were collected from various internet resources and existing WCE datasets. A comprehensive benchmarking and evaluation of the developed dataset was done using nine classification-based, three detection-based, and three segmentation-based deep learning models. The dataset is of high-quality, is class-balanced and contains single and multiple bleeding sites. Overall, our standard benchmark results show that Visual Geometric Group (VGG) 19, You Only Look Once version 8 nano (YOLOv8n), and Link network (Linknet) performed best in automatic classification, detection, and segmentation-based evaluations, respectively. Automatic bleeding diagnosis is crucial for WCE video interpretations. This diverse dataset will aid in developing of real-time, multi-task learning-based innovative solutions for automatic bleeding diagnosis in WCE. The dataset and code are publicly available at https://zenodo.org/records/10156571 and https://github.com/misahub2023/Benchmarking-Codes-of-the-WCEBleedGen-dataset.

Motivation & Objective

• To address the lack of a medically annotated, class-balanced WCE dataset specifically for bleeding classification, detection, and segmentation.
• To develop a comprehensive benchmark using diverse deep learning models across classification, detection, and segmentation tasks.
• To establish a standardized evaluation framework for future AI research in automatic WCE bleeding analysis.
• To support the development of real-time, multi-task learning systems for clinical deployment in gastrointestinal bleeding diagnosis.
• To enable reproducible research by providing publicly available dataset and code.

Proposed method

• The WCEbleedGen dataset was constructed from diverse sources, including internet repositories and existing WCE datasets, ensuring image diversity and real-world representation.
• The dataset contains 2,618 frames with balanced classes: 1,309 bleeding and 1,309 non-bleeding frames, each annotated with class labels, bounding boxes, and binary masks.
• Nine deep learning models were evaluated for classification, three for object detection, and three for semantic segmentation using standard metrics such as accuracy, precision, recall, IoU, and Dice coefficient.
• The benchmarking process included training and validation on the dataset, with performance evaluated on both validation and test sets to assess generalization.
• The models were trained and evaluated using standard deep learning frameworks, with hyperparameters tuned to optimize performance across tasks.
• The dataset supports multi-task learning by including combined annotations (class labels, bounding boxes, and segmentation masks), enabling joint learning across tasks.

Experimental results

Research questions

• RQ1 What is the performance of state-of-the-art deep learning models in classifying bleeding and non-bleeding frames in WCE using a newly curated, medically annotated dataset?
• RQ2 How do different detection models perform in localizing bleeding regions within WCE frames, and which architecture achieves the highest localization accuracy?
• RQ3 Which segmentation model achieves the best overlap with ground truth bleeding regions, and how does it generalize across diverse bleeding appearances and intensities?
• RQ4 Can the proposed dataset serve as a reliable benchmark for future AI models in WCE bleeding analysis, and how does it compare to existing datasets in terms of quality and annotation completeness?
• RQ5 What are the key limitations of current AI models in bleeding detection, and how can they be improved through better architecture design or data augmentation?

Key findings

• VGG19 achieved the highest performance in bleeding classification with strong accuracy, precision, and recall, indicating its effectiveness for frame-level binary classification.
• YOLOv8n outperformed other detection models, achieving the best balance of precision and recall in localizing bleeding regions across frames.
• LinkNet demonstrated the best segmentation performance, achieving high IoU and Dice coefficient values, indicating superior boundary localization and region overlap.
• SegNet showed high pixel-wise classification accuracy but poor boundary definition, resulting in lower IoU and Dice scores due to inaccurate segmentation boundaries.
• The benchmark results revealed that while classification and detection metrics were moderate, segmentation performance was strong, especially for LinkNet, highlighting its potential for clinical deployment.
• The dataset and code have been actively used in the Auto-WCEBleedGen challenges, confirming its utility and reproducibility in the research community.