Template-based Identification of Malignant and Non-malignant Skin-lesions to Minimize Biopsy Load Using Saturation Counts (HSV space) and Absolute Dark Area parameters: Experiments on ISIC Images

Abstract

Malignant skin lesion is the deadliest skin disease resulting in huge loss of lives in Europe, Australia and USA. Early detection of malignant lesions can save lives. It is highly challenging to differentiate between malignant and non-malignant skin lesions. Many non-invasive techniques have been proposed but none has been accepted in clinical practice. Consequently, biopsy remains the only gold standard for diagnosis of malignant lesions. The objective of this study which uses two master templates (MT) of dermoscopic images for identification, an improvement over the recently reported subtraction technique using only a single MT, is to propose a non- invasive technique to minimize biopsy load to an appreciable extent. This study proposes selection of two MTs, one a known 100^&#x0025; malignant (M) lesion, and the other a known nearly 100^&#x0025; benign (B) lesion. For identification of test lesions either belonging to <tex>$M$</tex> or B category, each test image from the publicly available ISIC archive is subtracted from each of the two MTs and the resulting pixels (RGB) data on each subtraction are converted into HSV space. Scatter plot showing Saturation (S) data counts against pixels locations below and above a trial-and-error-based threshold of 0.35, decides the B or M category of test lesions according to a rule defined for identification. The proposed method introduces, for the first time ever, use of double MTs subtraction technique, which amounts to the filter action. The proposed subtraction method has sound mathematical and logical base. On a preliminary trial over fifty images from publicly available ISIC archive, an overall high accuracy of 94&#x0025; was achieved which promises clinical applications to minimize biopsy load to a great extent. The proposed method is easy to implement by non-experts and takes only fifteen minutes on average for diagnosis.