Lesion classification using skin patterning

Abstract

BACKGROUND/AIMS: The observation that skin line patterning tends to be disrupted by malignant but not non-malignant skin lesions suggests that this could be used as an aid to lesion differentiation. Since recognised differentiating features can be obtained from the simply-captured white light optical image, the possibility of using such images for skin pattern disruption detection is worth exploring. METHODS: The skin pattern has been extracted from optical images by high-pass filtering and profiles of local line strength variation with the angle estimated using a new consistent high-value profiling technique. The resultant profile images have been analysed using a novel region-based agglomerative clustering technique (mRAC) and also a local variance measurement. A measure based on the relationship between the classification results and an intensity-based segmentation was calculated, and this represented the disruption of the skin line patterning. RESULTS: A set of images containing a variety of histologically confirmed malignant and non-malignant lesions was analysed. The computed textural disruption figure was compared to both the histological diagnosis and to a visual estimate of patterning disruption for each image. It was demonstrated that lesion separation could be achieved by both analysis methods, with a good correlation with visual estimate of disruption and with mRAC providing the best performance. CONCLUSIONS: It was concluded that the acquisition and modelling of skin line patterning from clinical images of skin lesions had been successfully achieved and that the analysis of the resulting data provided an assessment of pattern disruption that is both consistent with visual inspection and effective in presenting information useful for discrimination between melanoma and benign naevi lesion examples.