Classification of pathologicalshapes using convexitymeasures
Author: Paul L Rosin
DOI: 10.1016/j.patrec.2008.12.001
Abstract:
Two new shapemeasures for quantifying the degree of convexity are described. When applied to assessment of skin lesions they are shown to be an effective indicator of malignancy, outperforming Lee et al’s. OII scale–space based irregularity measure. In addition, the new measures were applied to the classification of mammographic masses and lung field boundaries and were shown to perform well relative to a large set of common shapemeasures that appear in the literature such as moments, compactness, symmetry, etc.
In image-based computer-aided diagnosis of suspected pathologies, classification is commonly determined by their colour, density, texture, morphology, etc. This paper focuses on the last characteristic, namely outline shape. Ideally, a shapemeasure should be non-parametric (i.e. free from tuning parameters), simple and efficient to implement and compute, robust, and invariant to transformations such as rotation, translation, and scaling. The starting point for the work described here was the paper by Lee et al. (2003) on developing a measure of irregularity which they applied to skin lesions in order to differentiate benign melanocytic nevi from malignant melanomas. They worked with an extensive set of 40 lesion borders with extensive ground-truth. each of which was assessed by fourteen dermatologists on a four point scale. Fig. 1 shows the skin lesion data as originally presented in (Lee et al., 2003), but reordered according to each lesion’s mean ground-truth score. While Lee et al. demonstrated that irregularity was a reasonable indicator of malignancy, examination of Fig. 1 also suggests that convexity is a strong factor.