Abstract
Wireless Capsule Endoscopy (WCE) is a non-invasive diagnostic tool for small intestine examination, yet it produces a massive volume of redundant images (60,000–120,000 per procedure), leading to significant diagnostic fatigue and potential oversight of lesions. This paper proposes a two-stage image screening framework to reduce redundancy. The first stage utilizes agglomerative hierarchical clustering based on global color profiles. The second stage identifies visually salient sub-image blocks and applies Structural Similarity Index (SSIM) analysis to these localized regions. Experimental results on 52 clinical cases demonstrate that this method can reduce diagnostic time from 78 to 18 minutes, achieving 100% recall of abnormalities at a 66.72% reduction ratio.
Introduction
The primary challenge in WCE is the "needle in a haystack" problem. Gastroenterologists must spend 2–3 hours reviewing a single study, where the vast majority of frames depict healthy tissue. Redundancy arises from the slow movement of the capsule. While automated classification via deep learning is an active area of research, these "black box" models often lack the interpretability required for clinical trust. This study addresses the redundancy problem by developing a transparent, two-stage filtering system that prioritizes localized structural changes in salient areas.
Methodology
The proposed screening method operates in two distinct phases:
1. Agglomerative Hierarchical Clustering
Initially, images are grouped by global color characteristics. Images are converted to the HSV (Hue, Saturation, Value) color space, and a 1,000-dimensional histogram vector is generated for each. An agglomerative hierarchical clustering algorithm then constructs a distance tree to group similar frames, effectively partitioning the video into clusters of visually related content.
2. Saliency-SSIM Keyframe Extraction
Within each cluster, the system identifies "visually salient" areas—regions that contrast significantly with the background, often indicating potential lesions. Using the Lab color space and Gaussian filtering, a saliency map is produced. The algorithm then selects the top three most salient 40x40 pixel blocks.
Instead of comparing entire frames, which can be dominated by static healthy tissue, the system calculates the SSIM for these specific salient blocks across adjacent frames. If the structural similarity is high, the frame is discarded as redundant. If the similarity is low, indicating a change in a potentially critical area, the frame is retained as a keyframe.
Experimental Results
The system was evaluated using a dataset of 68,400 images from 52 clinical cases, featuring 17 types of lesions. Performance was benchmarked against the manual clustering results of five independent gastroenterologists.
Efficiency and Safety
The implementation of this screening method reduced the average review time from 78 minutes to 18 minutes. At a reduction ratio of 66.72%, the system achieved 100% recall for all abnormal categories. When the reduction ratio was increased to 79.09% (matching human performance), the recall remained high at 94.23%, with missed items being clinically minor abnormalities.
Conclusion
By combining global clustering with localized saliency-based structural analysis, this method effectively mitigates the data bottleneck in WCE. The approach preserves the clinician's role as the final diagnostic authority while significantly reducing the cognitive load and time required for review. Future work will focus on refining feature extraction to capture even the most subtle lesions at higher reduction ratios.
Summary updated