New descriptor better distinguishes melanoma from benign lesions

A new descriptor, based on a lesion’s number of colors, uses the novel “number of colors difference,” or NCD, attribute instead of number of colors to better differentiate melanoma from benign lesions, according to a study published August 31 in the journal Medical and Biological Engineering and Computing.

The authors noted that analysis of skin lesions’ color is essential for accurately diagnosing melanoma. And the “number of colors” detected has been the most common and strongest melanoma indicator.

RELATED: Emerging options for high-risk melanoma​

Those calculating the number of colors in a lesion have to determine a color counting threshold. There is, however, no specific proposed method for determining color counting threshold in the literature, according to the paper.

Authors Hasan Akan, M.Sc., who lectures on electronics at Duzce University Vocational School, and Mustafa Zahid, Ph.D., a faculty member at Sakarya University of Applied Sciences, both in Turkey, did the study to define a new descriptor for diagnosing melanoma. They studied use of the new descriptor in three dermoscopic databases totaling hundreds of melanoma and benign samples.

Color counting thresholds notably affect the number of colors in study results. Research suggests the number of colors in melanoma are higher than in benign lesions. But these authors found that using the “number of colors” attribute can give opposite results, depending on the selected color counting threshold value. They also found color counting threshold values used in published studies were not statistically suitable in different data sets, and the most appropriate color counting threshold values for examining melanoma were 0 and for benign lesions 0.123.

Their finding that the 0 color counting threshold value was prominent in all the datasets they studied suggests that the smallest of color changes in a lesion should be taken into account when diagnosing melanoma, according to the paper.

Threshold value section alone, however, isn’t enough to improve classification performance.

That’s where NCD comes in, they write.

“NCD is not a simple color number attribute but a measure of the variation in the number of colors of the lesion for different [color counting thresholds]. In this respect it differs from the classic [number of colors] attribute,” the authors write.

RELATED: AI has potential to improve melanoma detection​

The dynamic range, which is 1 to 6 for number of colors attribute, expanded to 0 and 30 with NCD, allowing for the evaluation of even the smallest differences among lesions.

Univariate classification studies using NCD showed it was an important indicator for diagnosing melanoma. The average true positive for melanoma was up to 80% and the average true negative as high as 68.5% in their experiments. They wrote these values improve test accuracy, or f-measure, value as much as 52.7% and true positive value up to 84.5% when compared with results by the number of colors attribute that depends on different color counting thresholds, which researchers have used in published studies.

“The proposed [color counting threshold] determination method and the associated NCD calculation method can be easily applied to lesion images created under different conditions,” they add.

Using the new NCD feature in the ABCD system, which is commonly used in dermatology and is based on the criteria asymmetry (A), border (B), color (C), and differential structure (D), could be more accurate with the new descriptor, according to the paper.

Reference:
Akan, H., Yıldız, M.Z. Development of new descriptor for melanoma detection on dermoscopic images. Med Biol Eng Comput (2020). https://doi.org/10.1007/s11517-020-02248-z