|
Abstract:
|
Melanoma is the deadliest form of skin cancer, resulting in approximately 10,000 deaths annually in the United States. Early diagnosis by dermatologists helps decreasing mortality rates and healthcare costs. Smartphone applications employing computational tools to detect melanoma from images could improve early diagnosis. However, the best commercially available computational tool only correctly identifies melanoma lesions 72% of the time (true-positive rate) and benign lesions 36% of the time (true-negative rate), impeding diagnosis for 28% of patients. This study developed a novel ML (Machine Learning) system to select red-green-blue features from two dermoscopic datasets of pigmented lesions and implement an ensemble of five ML algorithms to detect melanoma based on images alone. Sixty-one melanoma detecting features were developed, surveying for lesion boundary regularity and brightness, image brightness distribution, and lesion color variation. Forty-one features were statistically significant (p< .05) for the first dataset, and thirty-five features were statistically significant for the second dataset. The ML model yielded a true-positive rate of 98% for the first dataset.
|
