Automated classification of malignant melanoma based on detection of atypical pigment network in dermoscopy images of skin lesions

“Melanoma causes more deaths than any other form of skin cancer. Early melanoma detection is important to prevent progression to a more deadly stage. Automated computer-based identification of melanoma from dermoscopic images of skin lesions is the most efficient method in early diagnosis. An automated melanoma identification system must include multiple steps, involving lesion segmentation, feature extraction, feature combination and classification. In this research, a classifier-based approach for automatically selecting a lesion border mask for segmentation of dermoscopic skin lesion images is presented. A logistic regression based model selects a single lesion border mask from multiple border masks generated by multiple lesion segmentation algorithms. This research also presents a method of segmenting atypical pigment network (APN) based on variance in the red plane in the lesion area of a dermoscopic image. Features extracted from APN regions are used in automated classification of melanoma. The automated identification of melanoma is further improved by fusion of other features relevant to melanoma detection. This research uses clinical features, APN features, median split cluster features, pink area features, white area features and salient point features in various hierarchical combinations to improve the overall performance in melanoma identification. A training set of 837 dermoscopic skin lesion images together with a disjoint test set of 804 dermoscopic skin lesion images are used in this research to produce the experimental findings”--Abstract, page iv