Aims:
Poor visualization of the small bowel due to the presence of intestinal content remains
one of the main limitations in capsule endoscopy (CE) procedures. The aim of our study
was to develop a tool that can automatically detect intestinal content in CE procedures.
Methods:
We created computer algorithms capable of distinguishing automatically between dirty
and clean regions in frames from CE videos. We extracted 563 frame images from 35
different CE videos. Each frame was divided in segments of 64 × 64 pixels, referred
to as patches. A total of 55293 patches were annotated by an experienced reader. We
assigned the frame images to two different sets: 80% for the training set and 20%
for the testing set. We extracted features based on colour and texture for discrimination
between clean regions and regions with intestinal content. With frames used for test
purposes we calculated accuracy (ACC), sensibility (S) and specificity (SP) in five
different models to analyze their performance. We then used the model to predict whether
the region is clean or contains intestinal content and also the pixel probability.
Results:
51,04% patches were classified as dirty regions and 48,96% as clean regions. We performed
5 different validation tests to evaluate different algorithms and their performance
in predicting a patch as either clean or dirty. We obtained an average accuracy of
87,12%, sensitivity of 89,89% and specificity of 84,50% using Supporting Vector Machine
(SVM) classification.
Conclusions:
Using patch probabilities, Endoclean system allows the estimation at a pixel level
of the percentage of cleanliness in images of CE videos with high accuracy. With optimization
of our results, this tool can be implemented for objective assessment of the quality
of mucosal visualization in CE procedures and can later provide the opportunity to
compare different types of preparations that can be used to improve the procedure
reliability.
