Measuring the Changes of Skin Lesions

Lesion Detection, Lesion Matching and Lesion Change Measurement

Ran Liu, Carlo Tomasi

In an attempt to build software tools to help dermatologists in measuring and detecting changes of lesions overtime, which is critical in early diagnosis of melanoma, we propose a framework consisting of three parts: lesion detection, lesion matching and lesion change measurement. In this project, we apply a locally adapted version of the Laplacian of Gaussian (LoG) filter to the green band of the skin images and a threshold of the filtering response is determined dynamically to discriminate lesion pixels. By formalizing the lesion matching problem as a generalized point pattern problem, a variant of the Iterative Closest Point Algorithm (ICP) called Fractional ICP (FICP) algorithm is proposed to find the correspondence between the two detected lesion clouds and the transformation aligning them. FICP algorithm is robust to outliers and is able to estimate the fraction of outliers automatically. Scale constraint affine transformation is used to align two lesion patterns and the difference after alignment is used as the measure to see how good our method can be in aligning unchanged lesions. We also discuss the possibilities of reducing the difference by using soft boundary images and adapted image blurring. Experiments are conducted on pairs of skin images collected by dermatologists overtime. It is shown that the algorithms given above are effective in detecting lesions, aligning them together and detecting lesion changes of clinical interest.