Aims Video capsule endoscopy is a minimally invasive tool for examining the small intestine,
though its adoption is limited due to intensive manual interpretation and battery
constraints. This study seeks to improve capsule endoscopy efficiency through edge
Artificial Intelligence (AI), which processes data on the device itself, enabling
real-time diagnostics at the point of care. The Galar dataset — a comprehensive, multi-center,
multi-label video capsule dataset — addresses these limitations in automated detection
and anatomical localization, bridging a gap as no edge AI solutions currently exist
for capsule endoscopy. This project facilitates broader clinical use and potential
improvements in patient outcomes.
Methods The Galar dataset, compiled across Germany, contains over 3.5 million frames from
80 capsule videos annotated for technical, anatomical, and pathological features.
Annotations by five expert annotators used the Computer Vision Annotation Tool (CVAT),
with accuracy validated via ResNet-50 model training. Based on our dataset, our collaborative
research focuses on developing edge AI approaches that improve model performance.
First, Convolutional Neural Networks (CNNs) were combined with Hidden Markov Models
(HMMs) for time-series analysis, enabling accurate localization within the gastrointestinal
tract while employing a low-parameter model (approx. 1 million parameters) fit for
low-power devices. Additionally, ensemble models, combining image classifiers and
autoencoders, optimized detection accuracy while minimizing computational load.
Results The Galar dataset stands as one of the most comprehensive resources in capsule endoscopy,
with 29 anatomical and pathological labels essential for gastrointestinal diagnostics.
The collaborative edge AI approach, integrating CNNs with HMMs, achieved a 93% accuracy
on the dataset, demonstrating precise anatomical localization. Ensemble models also
reached high accuracy in distinguishing features within the dataset, achieving an
AUC score of 76% for anomaly detection. These low-complexity models align with capsule
endoscopy's energy constraints, enabling real-time, autonomous evaluations on resource-limited
devices. Comparatively, this edge AI model achieves similar, if not improved, outcomes
with fewer resources than traditional CNN models.
Conclusions As one of the largest annotated public datasets for video capsule endoscopy, the
Galar dataset is a valuable resource for AI advancements in gastrointestinal imaging.
Findings indicate that low-parameter, energy-efficient models enabled by edge AI can
enhance diagnostic efficiency in capsule endoscopy, lessen clinician workload, and
enable real-time anomaly detection directly on the capsule. These models have the
potential to perform preliminary analyses on-device, sending only relevant images
or real-time capsule localization and conserving battery life. This advancement positions
video capsule endoscopy as a more efficient, accessible diagnostic tool.
