Classification of videocapsule endoscopy image patterns: comparative analysis between patients with celiac disease and normal individuals

Abstract

Background: Quantitative disease markers were developed to assess videocapsule images acquired from celiac disease patients with villous atrophy, and from control patients.

Method: Capsule endoscopy videoclip images (576 x 576 pixels) were acquired at 2/second frame rate (11 celiacs, 10 controls) at regions: 1. bulb, 2. duodenum, 3. jejunum, 4. ileum and 5. distal ileum. Each of 200 images per videoclip (= 100s) were subdivided into 10 x 10 pixel subimages for which mean grayscale brightness level and its standard deviation (texture) were calculated. Pooled subimage values were grouped into low, intermediate, and high texture bands, and mean brightness, texture, and number of subimages in each band (nine features in all) were used for quantifying regions 1-5, and to determine the three best features for threshold and incremental learning classification. Classifiers were developed using 6 celiac and 5 control patients' data as exemplars, and tested on 5 celiacs and 5 controls.

Results: Pooled from all regions, the threshold classifier had 80% sensitivity and 96% specificity and the incremental classifier had 88% sensitivity and 80% specificity for predicting celiac versus control videoclips in the test set. Trends of increasing texture from regions 1 to 5 occurred in the low and high texture bands in celiacs, and the number of subimages in the low texture band diminished (r(2) > 0.5). No trends occurred in controls.

Conclusions: Celiac videocapsule images have textural properties that vary linearly along the small intestine. Quantitative markers can assist in screening for celiac disease and localize extent and degree of pathology throughout the small intestine.

Figure 1

Plot of surface variability (standard deviation) versus mean surface brightness (units are grayscale level from 1 - 256) from 10 × 10 pixel subimages in one control (left) and one celiac (right) videoclip image frame. Ranges of 10 grayscale units in standard deviation of pixel brightness are called bands. Many of the subimage values are clustered in band 1 (below 10 grayscale level in standard deviation). However, many celiac subimage values are scattered at higher bands, meaning that there is more variation (higher standard deviation) in the celiac subimages.

Figure 2

Development of the threshold classifier to distinguish celiac from control videoclips using exemplar data. The threshold level T (vertical black line) was determined so as to maximize the sum of the celiac and control videoclips correctly classified (noted by the x). For example in the bottom panel, the threshold is set at -0.7 normalized grayscale units. This threshold level results in the correct classification of all but 3 celiac videoclips (to left of line in upper histogram) and all but 7 control videoclips (to right of line in lower histogram). Of the nine features that were measured (see Methods) the three features shown provided the best threshold classification.

Figure 3

A. Use of the three-dimensional threshold classifier on a set of exemplar videoclips, with best projection and best nonlinear discriminant function for classification shown. The features used for classification and orientation of the coordinate axes are noted at top. Most of the control videoclip scatterpoints (red) reside at lower right i.e. approximately aligned with the x-axis, with low degree of texture in the 0-10 band. The nonlinear discriminant function (white curved line) mostly separates the two groups correctly. Only 1 control scatterpoint (red) is clustered with the celiac scatterpoints (blue), and only 2 celiac scatterpoints are clustered with the control scatterpoints (panel A). B. Use of the incremental learning classifier on a set of exemplar videoclips, with best projection and best nonlinear discriminant function for classification shown. The features used for classification and orientation of the coordinate axes are again noted at top.

Figure 4

Classification of a test set of videoclips using A. Threshold classifier and B. Incremental learning classifier. Nonlinear discriminant functions are shown as white curved lines. The features used for classification and orientation of the coordinate axes for both classifiers are noted at top. Using the same orientation and discriminant functions as for the exemplar set (Figure 3) most celiac (blue) and control (red) videoclip values are correctly classified.

Figure 5

Mean values of each of the nine features at each small intestinal level were computed from the pooled videoclips of all celiacs, and the resulting scatterplots and linear regression lines are shown. The change along the small intestine was considered to be a trend for coefficient of determination r² > 0.50. A trend is therefore defined as an approximately linear change in feature value from location 1 to 5 (i.e. from proximal to distal along the small intestine).

Figure 6

Mean values of each of the nine features at each small intestinal level were computed from the pooled videoclips of all controls, and the resulting scatterplots and regression lines are shown. None of these were considered to have significant trends from duodenal bulb to distal ileum (r² ≤ 0.50). The lack of trend in image brightness and texture in the control population suggests a lack of pathologic changes from proximal to distal along the small intestine.