Supplement Title Quantitative Laryngoscopy With Computeraided Diagnostic System For

supplement: title: quantitative laryngoscopy with computer-aided diagnostic system for laryngeal lesions chung feng jeffrey kuo1#, wen-s

Supplement:
Title: Quantitative laryngoscopy with computer-aided diagnostic system
for laryngeal lesions
Chung Feng Jeffrey Kuo1#, Wen-Sen Lai2#, Shao-Cheng Liu3*
1Department of Materials Science & Engineering, National Taiwan
University of Science and Technology, Taipei, Taiwan, Republic of
China
2Department of Otolaryngology-Head and Neck Surgery, Taichung Armed
Forces General Hospital, Taichung, Taiwan, Republic of China
3Department of Otolaryngology-Head and Neck Surgery Tri-Service
General Hospital, National Defense Medical Center Taipei, Taiwan,
Republic of China
#Co-first authors
＊Corresponding author:
Shao-Cheng Liu, M.D. PhD.
Associate Professor
Department of Otolaryngology-Head and Neck Surgery,
Tri-Service General Hospital, National Defense Medical Center,
No. 325, Sec. 2, Cheng-Gong Road, Neihu District, Taipei, Taiwan 114,
R.O.C.
Tel: 886-2-8792-7192
Fax: 886-2-8792-7193
E-mail: [email protected]
Running title: Computer-aided quantitative laryngoscopy
Author contribution:
1. Chung-Feng Jeffrey Kuo, PhD: study design, critical article
review/editing.
2. Wen-Sen Lai: acquisition of data, data analysis and interpretation,
article and images review/editing.
3. Shao-Cheng Liu, MD, PhD: study design, data collection, literature
search, images editing, article drafting, article submission.
Supplement:
Contrast-limited adaptive histogram equalization (CLAHE)
There exist some problems such as uneven brightness distribution or
low contrast in laryngoscope images, which will lead to inaccurate
segmentation and inconsistent eigenvalues. In this study, CLAHE was
used to improve [6-8].
First, a global histogram equalization was performed on an image of
the laryngoscope, resulting in the loss of details in areas where the
images were too bright or too dark. Therefore, an adaptive histogram
equalization (AHE) with good local contrast was used to divide the
images into different grid regions, which limited the histogram to a
small region so as to perform operations via different regions.
Expanding the grayscale values of the upper and lower limits from 0 to
255 by contrast enhancement of the maximum and minimum gray values of
the region; if the region was large, the contrast would be reduced,
the area was small, and the contrast would be enhanced, which can
improve the situations of overall strong or dull image contrast, or
partial loss of details. When the pixels contained in the region were
very similar, the histogram would be very concentrated, causing a very
narrow range of pixels to be mapped to the entire pixel range,
followed by amplifying the noise of the local region at the same time.
To solve the problem of noise amplification, contrast-limited adaptive
histogram equalization (CLAHE) used the contrast limited method to
equally distribute the histogram above the set height to the low-end
histogram, thereby reducing the slope of the cumulative distribution
function, and the higher the threshold setting was, the higher the
contrast would be. Later, AHE could be used to effectively solve the
additional problem of noise amplification due to the contrast
adjustment.
Image smoothing
Swaying or swallowing, saliva reflecting light of laryngoscope images
leads to images noises, which needs to be smoothed to remove noise. In
this study, Gauss smoothing was used to change the core parameters,
improve the areas where the gray scale values change dramatically,
such as saliva reflecting light etc., to reduce the segmentation
errors in subsequent steps. Gaussian smoothing increased the weight of
the pixel near the center, which reduced the fuzzy phenomenon after
processing. The parameters of the Gaussian template were calculated by
the Gaussian function, and the current pixel values were replaced by
adding the result of each pixel in the template, and then the
convolution was used to process the entire image. The Gaussian
function is shown as Eq. (1)

(1)
where is the standard deviation and , are the
image pixel locations. If the standard deviation is too small,
the off-center pixel weight will be very small, and the result is like
there is no processing. If the standard deviation is too
large, the Gaussian template will degenerate into an average template.
Actually, when the template is 3×3, is 0.8. For a larger
template, the value of can be appropriately increased. The
Gaussian template is shown in S-Tab. 1.
1/16
2/16
1/16
2/16
4/16
2/16
1/16
2/16
1/16
S-Tab. 1. The Gaussian template
The Fast Otsu method
The glottis is the central area of the throat. Segmenting the glottis
is conducive to the subsequent vocal cord segmentation. The glottis
belongs to the dark area in the image, so the grayscale value was used
to segment the glottis. In order to highlight the difference of image
brightness and retain more details, and avoid the result of
over-segmentation, Fast Otsu method [9], which is less computational,
was used to improve the disadvantage that the traditional Otsu method
needs to calculate the inter-group variance of all grayscale values in
order to obtain the best threshold value. The Fast Otsu method
searches for the trend of the threshold value by calculating the
variance between groups of the grayscale value squares, estimates the
area where the optimal threshold value is located, and narrows the
search region by iterating continuously, to find the grayscale value
corresponding to the maximum intergroup variance, that is, the optimal
threshold value.
The steps of the Fast Otsu method are as follows:
1. Calculating the average grayscale value of the entire image (
), images above the average grayscale value are regarded as the
foreground, and vice versa as the background, and calculating the
average grayscale value of the two.

(2)

(3)

(4)
where is the average background value, is the average
foreground value, is the grayscale value 1~ , and
is the probability of the grayscale value
2. According to step 1, the search area is divided into 0~ ,
~ , ~ and ~255. The inter-group
variances of the four regions were calculated separately, and the
largest inter-group variance was regarded as the possible optimal
threshold.
3. Calculating the inter-group variances of the previous grayscale
value and the next grayscale value of the possible optimal threshold
value respectively. If the former value is the largest, the search
direction proceeds to the previous region, and if the latter value is
the largest, the search direction proceeds to the next region. If the
middle value is the largest, the current threshold value is the
optimal threshold value, and the iteration stops.
4. After the search direction is known by step 3, the new area
~ is re-established, and the area is divided into 4 small equal
parts, of which three dividing points of the 4 equal parts are
calculated as follows.

(5)

(6)

(7)
Calculating the inter-group variances of the grayscale values of the
three dividing points, regard the largest inter-group variance as the
potential optimal threshold and return to step 3 for confirmation.
ACM
The vocal cords are blurred with the surrounding tissue boundaries. In
this study, ACM [10] was used as the image segmentation method. The
algorithm defines the initial parameterized curve, and makes the curve
move to the target boundary by minimizing the energy function, i.e.
the reconciliation of the internal and external forces of the image,
and the boundary extraction is completed when the energy function
reaches the minimum,
First, in the energy function, represents the parameter curve,
, and represent the and
coordinates on the target contour, respectively. The image is
represented by , and the energy function is Eq. (8)

(8)
where , are positive real numbers greater than 0, the
first and second terms of the energy function are collectively
referred to as internal energy (internal force) for the purpose of
maintaining the continuity and stability of the curve, and the first
one is theㄧorder differential term of the contour, which resists the
tensile force mainly by controlling the slope of the contour; the
second one is the second order differential term of the contour, which
is used to resists the bending moment force.
The third term of the energy function is external energy, and
is defined as Eq. (9).

(9)
The external force of the image is composed of three forces,
is the individual weight of the three forces. represents the
line energy of the image , as Eq. (10)

(10)
where is the Gaussian standard deviation, and is the
image matrix.
represents the edge energy of the image as Eq. (11)

(11)
denotes terminations energy. To obtain the termination
condition, Kass et al. defined the curvature of the horizontal line on
Gaussian blurred image. Let be the Gaussian blurred image
, define the gradient angle, and and
represent the unit vectors horizontal and perpendicular to the
gradient direction, respectively.
The horizontal contour curvature is shown as Eq. (12).

(12)
The initial parameterized curve is mutually pulled by the internal and
external forces of the energy function mentioned above, and finally
the balance of the forces, that is, the minimized energy function is
achieved, and the final contour boundary is obtained.
Supplementary Figure 1. System image processing flow

Supplementary Figure 2. Filter clear vocal cord images. (a)The glottis
area is the largest, and the vocal cord image is clear. (b)The glottal
area is too small, and the vocal cord image is blurred, so that the
vocal cord condition cannot be easily identified.

Supplementary Figure 3. Image shielding.
The peripheral areas in the original image (a) are black by the naked
eye observation, but in computer analysis, these peripheral areas are
composed of pixels with low-value grayscale pixels, which will
interfere with subsequent segmentation and features analysis. In this
study, the peripheral grayscale pixels were unified to 0 (b), so that
the accurate values can be obtained by excluding the grayscale pixels
of 0 from the subsequent calculations.

Supplementary Figure 4. Vocal cord seed point process. (a) Glottic
image. (b) Glottic contour. (c) Vocal cords seed point.

Supplementary Figure 5. Vocal cord lesion features
The structure of the cyst is liquid and the shape is flat.
The structure of polyps is soft meat and the shape is convex.
直線單箭頭接點 711 直線單箭頭接點 712

Fig.5. Polyp and cyst shape difference