Table 1.
Mayo endoscopic subscore (MES).
Fig 1.
A, An extracted still picture with the size of 576 × 576 pixels. The patch dataset images (128 × 128 pixels) were trimmed from the still picture starting from the left upper corner (white dotted patch), rightwards (white solid patch), then downwards (red solid patch) at every 32-pixel-strides (white and red arrows) over the entire effective region of the still picture. B, A total of 40 patches eligible for analysis.
Fig 2.
Representative images of the six categories: 0) MES0, normal (upper), white scar (middle) and inflammatory polyps (lower); 1) MES1, decreased vascular pattern; 2) MES2, absent vascular pattern, friability, and erosions; 3) MES3, ulceration; 4) inadequate quality for evaluation, effluent with residue (upper), bubble (middle) and motion blur (lower); and 5) ileal mucosa.
Table 2.
Number of images in each training and validation dataset.
Fig 3.
Algorithm for evaluating the severity of a single still picture.
Patches trimmed from input images (left columns of A–D) were classified into MES0 (dark gray open square), MES1 (yellow open patch), MES2 (magenta open patch), and MES3 (red open patch). Area0 (area of MES0) is defined by the union of the dark gray open patches. Similarly, area1, area2, and area3 by that of yellow, magenta, and red open patches, respectively. Severity is expressed by the stacked bar graph, composed of % area: white, MES0; yellow, MES1; magenta, MES2; and red, MES3 (right columns of A–D).
Fig 4.
Training images of ulceration.
A and B, examples for training images of ulceration with exudate labeled as MES3; C and D, examples for training images with opaque residue labeled as inadequate quality for evaluation which were not discriminated from A and B; E and F, examples for validation images with minor ulceration labeled as MES3, which were misclassified as white scar or MES0.
Table 3.
Accuracy of the training and test data set.
The accuracy for each category is presented next to the number of images.
Table 4.
Confused matrix showing the classification results using established convolutional learning network.
Fig 5.
Examples of topography map of severity in the same patient along the entire length of the colorectum (the cecum and ascending, transverse, descending and sigmoid, and rectum) before (A) and after (B) therapeutic intervention. Suffix (-f) and (-b) indicate data files from forward and backward cameras, respectively. Severity is expressed by the stacked column composed of light gray (MES0%area), yellow (MES1%area), magenta (MES2%area), and red (MES3%area). The blank column corresponds to a still picture estimated as an inadequate condition for analysis.