Determining the Size of a Skin Lesion Using Smartphones
Jae Won Shin
University of Missouri
Trevor Ash
University of Mary Hardin-Baylor
Yi Shang
University of Missouri
Introduction
The objective of this project is to determine the area of a skin lesion from a photo and calculate the lesion diameter to help automatically diagnose melanoma. The American Cancer Society recognizes the
guidelines of asymmetry, border irregularity, color irregularity and diameter as markers for possible melanomas; lesions with a diameter greater than 6 mm may need to be examined by a doctor. Not all lesions are
circular, therefore the size will be determined by pixel area rather than diameter. Because of variable distances between the phone camera and the lesion in question, the area of the lesion changes in
comparison to the picture. Two methods have been developed for determining the area of a lesion within a close proximity to the camera.
Focus-distance-based Method
Reference-based Method
This method uses the focus distance returned by the getFocusDistances() method from Android API level 9 to
determine the distance from the camera to the lesion. Then from the pixel area of the lesion image, found through
the image processing capabilities of the OpenCV library, and the distance from the camera, the real area of the
lesion is found.
Using the average distance, that is (Near*Optimal*Far)/3, from
every trial and the pixel area of the coins found from those
trials a function was found to relate area in pixels2 to area in
millimeters2.
y = 13.885x2 - 455.42x + 4079.1
(where there are y pixels in a square millimeter at distance x)
Optimal
400000
250000
200000
11 CM
12 CM
150000
13 CM
Distance in centimeters
Actual Distance
12 CM
13 CM
Optimal
Near
Median
Far
11 CM
14 CM
15 CM
50000
0
0
1
2
3
4
5
6
7
8
9
10
1
2
3
MEAN
Actual
Nickel
Area
11 CM
0
50
100
150
200
250
300
350
400
450
Estimated area in square millimeters
Optimal
Average Estimated Quarter Area
Near
15 CM
Distance in Centimeters
Optimal
Near
0
10
20
30
40
50
60
70
Distance returned by camera
Focus-distance Normalizing Functions
Using the means of the near, optimal and far values
between 10 and 15 cm functions were found to
make the values returned from the method more
reliable
•  Near: x = (y+41.92)/5.3
•  Optimal: x = (y+57.51)/6.8714
•  Far: x = (y+85.591)/9.6131
5
6
7
8
9
10
Analysis
MEDIAN
12 CM
10 CM
Far
4
Trials
13 CM
Mean
Optimal
10 CM
13 CM
14 CM
Far
14 CM
13 CM
pix
MEDIAN
MEAN
12 CM
KRmm
Actual
Quarter
Area
11 CM
10 CM
380
400
420
440
460
480
500
520
Estimated area in square millimeters
mm2.
The “red flag” area for a lesion is around 28
Because most of the area estimations were off by over 28
mm2, a better estimator needed to be found.
Average Nickel Diameters (Reference: Quarter)
15 CM
Finding unknown area from known area
•  Found pixels per millimeter(ppm) by dividing
the known region area in pixels by its area in
millimeters KR
14 CM
pix
Finding Diameter from Area Estimates
ppm
It was found that the error was reduced greatly by
calculating diameter from estimated area.
13 CM
Median
12 CM
Mean
Actual
Nickel
Diameter
11 CM
10 CM
= ppm
20.7
20.8
20.9
21
21.1
21.2
21.3
21.4
21.5
Diameter in millimeters
Average Nickel Diameter Percent Differences
•  Then dividing the unknown region area in pixels
by ppm value to return unknown region area in
millimeters UR
1.40%
1.20%
1.00%
= URmm
0.80%
Mean
0.60%
Median
0.40%
0.20%
0.00%
10 CM
11 CM
12 CM
13 CM
14 CM
15 CM
Distance in centimeters
Average Nickel Area (Reference: Quarter)
(Where x is the real distance and y is the returned value)
Average Quarter Diameters (Reference:Nickel)
15 CM
Average Estimated Quarter Diameters
Average Estimated Nickel Diameters
Distance in centimeters
Far
15 CM
Optimal
15 CM
Near
Optimal
Near
Far
Optimal
Near
Median
Far
Optimal
14 CM
13 CM
Median
12 CM
Mean
11 CM
Actual
Nickel
Diameter
Mean
10 CM
Near
Distance in centimeters
Far
Distance in centimeters
14 CM
13 CM
Median
12 CM
19
Optimal
19.5
20
20.5
21
21.5
22
Median
13 CM
Mean
Actual
Nickel
Area
12 CM
10 CM
11 CM
10 CM
22.5
23
23.5
24
24.5
25
25.5
Actual
Quarter
Diameter
15 CM
14 CM
Mean
22
Far
14 CM
11 CM
13 CM
340
345
350
Area in
355
360
365
Mean
Actual
Quarter
Diameter
10 CM
millimeters2
23.7
Estimated diameter in millimeters
22.5
Median
12 CM
11 CM
335
23.8
23.9
24
Near
24.1
24.2
24.3
24.4
24.5
24.6
Diameter in millimeters
Estimated diameter in millimeters
Far
Optimal
Near
2
4
6
8
10
12
14
16
Average Quarter Diameter Percent
Differences
Normalized Returned Distances in centimeters
Average Percent Error
9.00%
8.00%
Average Nickel Diameter Percent
Differences
7.00%
8.00%
6.00%
7.00%
7.00%
Mean
3.00%
5.00%
1.20%
Mean
4.00%
Median
Median
3.00%
4.00%
Mean
3.00%
Median
15 CM
5.00%
4.00%
6.00%
1.40%
6.00%
5.00%
2.00%
2.00%
1.00%
14 CM
1.00%
13 CM
Median
12 CM
Mean
Actual
Quarter
Area
11 CM
1.00%
440
0.00%
10 CM
11 CM
12 CM
13 CM
14 CM
15 CM
0.80%
Mean
0.60%
Median
0.40%
10 CM
2.00%
1.00%
Average Quarter Diameter Percent Differences
Average Quarter Area (Reference: Nickel)
Percent Error
0
Distance in centimeters
15 CM
12 CM
200000
Trials
Near
10 CM
11 CM
250000
100000
15 CM
Optimal
Far
Real Distance in centimeters
11 CM
12 CM
13 CM
14 CM
10 CM
300000
150000
14 CM
50000
15 CM
Near
Percent Difference
350000
10 CM
Distance in centimeters
14 CM
Far
450000
300000
100000
Average Estimated Nickel Area
Near
500000
Distance in centimeters
15 CM
Far
350000
Percent Error
•  Multiple pictures taken of Nickel and Quarter
coins at each 1 cm interval from 10 to 15 cm
•  Output from getFocusDistances(): Near, Optimal,
and far were recorded
Finding the diameter from area
Multiple pictures taken of Quarter and Nickel coins
In
order
to
compare
results
between
the
two
in intervals of 1 cm from 10 cm to 15 cm
Nickel Pixel Area Readings
Quarter Pixel Area Readings
methods, the findings must be of the same
type. Because of this, the areas found from the
reference-based method were converted into
diameters despite the relatively small error
rates found using a reference.
Area in pixels2
Experiment
Experiment
Area in pixels2
Area from Focus-distance
This method uses uses a coin of known size as a reference and utilizes the automatic image segmentation and
recognition capabilities of OpenCV on Android to calculate the area of a lesion.
0.00%
10 CM
11 CM
12 CM
13 CM
14 CM
15 CM
445
450
455
460
Area in millimeters2
465
470
475
0.20%
0.00%
10 CM
0.00%
Near
Optimal
10 CM
Far
Near
Optimal
11 CM
Far
Near
Optimal
Far
Near
Optimal
12 CM
13 CM
Distance in centimeters
Far
Near
Optimal
14 CM
Far
Near
Optimal
15 CM
Far
11 CM
12 CM
13 CM
14 CM
15 CM
Distance in centimeters
Conclusions
Experimental results show that the reference-based method produces diameter estimations with errors typically less than 3% and an average error of 0.96%. The errors of the focus-distance-based method are
less than 13% with an average error of less than 5%. Though more accurate, the reference-based method requires the user to have a coin with them when they use the app. Both methods are successful and
are being incorporated into our automatic melanoma detection app on Android smartphones. Further work can be done in determining the symmetry or lack there of present in the lesion as another
criterion for classification.
Acknowledgements
MU Computer Science Department
NSF REU Program