Multispectral imaging approach in the diagnosis of cutaneous melanoma: potentiality and limits

Abstract

In an attempt to overcome the subjectiveness of clinical observation in the diagnosis of cutaneous melanoma, a computerized method is proposed. Reflectance images of 237 pigmented lesions (67 melanomas and 170 non-melanomas) were analysed using a telespectrophotometric technique. This device consists of a CCD camera with 17 interference filters. Images were acquired at selected wavelengths, from 420 to 1040 nm. Morphological and reflectance related parameters were extracted from the wavelength-dependent images of the lesions. The most significant features in the comparison between benign and malignant lesions were: lesion dimension (P < 10(-8) at 578 nm); mean value (P < 10(-7) at 940 nm) and standard deviation (P < 10(-4) at 904 nm) of lesion reflectance; lesion roundness (P < 10(-5) at 461 nm); and border irregularity (P < 10(-4) at 461 nm). Based on these parameters, a discriminant function between the two populations of lesions (naevi and melanomas) was obtained. By using the results of the analysis of the recruited lesions as 'training data', discriminant functions enabled the assignment of a score, or a 'risk probability', to each studied lesion. By imposing a sensitivity of 80% (a figure that mimics the diagnostic capability of an experienced clinician), entering or not entering the lesion dimension as input data in the discriminant analysis led to a specificity of 51% or 46% respectively. The high number of false-positive cases, which is a consequence of the selection criteria of the lesions, is, at present, the major limitation of the current technique. Nevertheless, our results suggest that an imaging-based computer-assisted device could be capable of discriminating malignant lesions mainly by evaluation of reflectance, especially in the infrared region, and shape properties. The dimension of a lesion should not be essential in the diagnosis of melanoma and, in our opinion, small melanomas should be recognized by a computer system as well as they are on clinical grounds.