Extraction of quantitative blur measures for circumscribed lesions in mammograms

This study investigates ways of improving lesion diagnosis in mammograms by deriving quantitative descriptions of the lesion periphery. The descriptions are derived by computer image analysis methods. The degree of blur at lesion boundaries is of prime concern, as poorly outlined lesions can indicate malignancy. The need for quantitative analysis arises from psychological evidence suggesting that the human visual system cannot precisely estimate the degree of blur. To help find suitable measures a set of 'artificial' lesions has been generated by convolving a step-like edge with a set of Gaussian functions G(sigma) where sigma characterizes the degree of blur. From these generated lesion images the parameters sigma are derived by the process involving deconvolution. As the edge changes are most important in radial directions, the measures of sigma are calculated for each radial profile of the lesion. The derived individual values correspond very closely to those used to generate the lesions. Statistical measures obtained from them allow distinction between edges which are blurred to different extents and yet are impossible to differentiate visually. The artificial lesions will be combined with mammographic data, and similar measures derived. The work will be validated on real lesions for which the histological findings are known from performed biopsies.