-Capsule endoscopy is a recent technology with a clear need for automatic tools that reduce the long exam annotation times of exams. We have previously developed a topographic segmentation method, which is now improved by using spatial and temporal position information. Two approaches are studied: using this information as a confidence measure for our previous segmentation method, and direct integrating of this data into the image classification process. These allow us not only to automatically know when we have obtained results with error magnitudes close to human errors, but also to reduce these automatic errors to much lower values. All the developed methods have been integrated in the CapView annotation software, currently used for clinical practice in hospitals responsible for over 250 capsule exams per year, and where we estimate that the two hour annotation times are reduced by around 15 minutes.
Index Terms— Endoscopic capsule, image classification, biomedical engineering, medical imaging
The clinical importance of the endoscopic capsule is now
solidly established in literature: Iddan [
The main contribution of this paper is the improvement of our previous topographic segmentation methods using color and texture, by incorporating not only temporal but also spatial position information in the image classification process.
The ultimate objective of the presented methods is to reliably divide the video of the gastrointestinal tract into its 4 constituent parts (entrance, stomach, small intestine, large intestine) and thus determine its corresponding junctions (esogastric junction, pylorus, ileo-cecal valve).
A. Capsule Position and Velocity
We can theoretically estimate the spatial position of a capsule via antenna signal triangulation. We have selected 47 capsule exams where a clinical specialist manually annotated the temporal location of the pylorus (tPYL) and of the ICV (tICV) in the video using the CapView annotation software. We’ve then used our automatic topographic segmentation algorithm to determine these same temporal locations. Using our 2D position information, we can then obtain the corresponding spatial locations: xPYL, yPYL, xVIC, yVIC etc. For comparison purposes, these were normalized. Besides analysing 2D position information, we have looked at average capsule displacement velocity (module of the displacement vector between two points with temporal references t and t+1).
B. Topographic Segmentation Algorithm
Our previously developed automatic topographic
segmentation method, from now on referred as TSA, is
described in Coimbra [
C. Spatial Information as a Confidence Measure
Two high-confidence areas were defined, one for the pylorus
and another for the ILC. We have measured the median
segmentation error SEz for all marks (z12 - eso-gastric
junction; z23 – pylorus; z34 – ileo-cecal valve), and for all
exams SE, whose junctions are inside and outside these areas,
D. Integrating Spatial and Temporal Information for
Classification
An alternative way of using this information is to use it
directly for individual image classification. Our previous
method trained 4 SVM classifiers, one for each zone, which
determine the topographic section each image belongs to as
the classifier with the highest positive distance to the SVM
hyperplane (see [
U
F > x, y, Z1 , Z 2 , Z 3 , Z 4 ,V , t@ (1)
where x and y are the normalized spatial location coordinates
(1,2), Z1, Z2 Z3 Z4 are the SVM classifier results [
A variety of well-known distances was used for classification (L1 Norm, Euclidean, Mahalanobis). Finally, we have measured the relevance of each coefficient for the segmentation process using a step-wise elimination analysis.
1 0.5 -1 Fig. 1. Spatial distribution of correct (green) and incorrect (blue) estimations. Points in high confidence areas are highlighted with a black bounding box. We can observe that most correct detections fall into high-confidence areas while incorrect ones are more distributed over the whole 2D space.
Accuracy 58 % 92 % Mean Err. 3966 2096
ICV and results presented in section 3 have showed that this information is indeed useful as a confidence measure for automatic segmentation results.
An analysis of Table 1 and Figure 1 shows that highconfidence areas contain almost all correct estimations, and low-confidence areas mainly contain incorrect estimations. 1 0.5 -0.5 y Z2 Z3 Z4 V 82.3% 81.8% 81.9% 80.9% 80.8% 83.2% 80.4% 82.3% 83.1% 81.2% 81.0% 82.6% 82.5% 82.4% 79.9%
Median Segmentation Error 83.5% 82.8% 82.2% 82.1%
79.2% 83.1% 83.1% 81.7% 82.8% 83.4% 83.5% 83.1% 79.9% 82.8% 82.8% 80.8% 82.8% 79.3% 82.2% 82.8% 79.5% 82.2% 82.2% 79.9% 82.2% 70.3% 82.2% 82.1% 66.4% 82.1% 82.1% 79.2% 82.1% 69.8% 82.1% 82.1% 63.0% 79.2% 79.2% 79.2% 79.2% 68.7% 79.2% 79.2% 54.2%
Results show that doctors can trust that automatic segmentation errors in high-confidence areas are as low as human ones. Including other information has allowed us to improve segmentation results significantly. Step-wise elimination analysis has shown us that the most relevant features for segmentation are capsule temporal position, and the color recognition of the entrance and the small intestine topographic sections. It has also shown us that spatial location is not a relevant factor for individual image classification.