Multispectral imaging and artificial neural network: mimicking the management decision of the clinician facing pigmented skin lesions

Various instruments based on acquisition and elaboration of images of pigmented skin lesions have been developed in an attempt to in vivo establish whether a lesion is a melanoma or not. Although encouraging, the response of these instruments, e.g. epiluminescence microscopy, reflectance spectrophotometry and fluorescence imaging, cannot currently replace the well-established diagnostic procedures. However, in place of the approach to instrumentally assess the diagnosis of the lesion, recent studies suggest that instruments should rather reproduce the assessment by an expert clinician of whether a lesion has to be excised or not. The aim of this study was to evaluate the performance of a spectrophotometric system to mimic such a decision. The study involved 1794 consecutively recruited patients with 1966 doubtful cutaneous pigmented lesions excised for histopathological diagnosis and 348 patients with 1940 non-excised lesions because clinically reassuring. Images of all these lesions were acquired in vivo with a multispectral imaging system. The data set was randomly divided into a train (802 reassuring and 1003 excision-needing lesions, including 139 melanomas), a verify (464 reassuring and 439 excision-needing lesions, including 72 melanomas) and a test set (674 reassuring and 524 excision-needing lesions, including 76 melanomas). An artificial neural network (ANN(1)) was set up to perform the classification of the lesions as excision-needing or reassuring, according to the expert clinicians' decision on how to manage each examined lesion. In the independent test set, the system was able to emulate the clinicians with a sensitivity of 88% and a specificity of 80%. Of the 462 correctly classified as excision-needing lesions, 72 (95%) were melanomas. No major variations in receiver operating characteristic curves were found between the test and the train/verify sets. On the same data set, a further artificial neural network (ANN(2)) was then architected to perform classification of the lesions as melanoma or non-melanoma, according to the histological diagnosis. Having set the sensitivity in recognizing melanoma to 95%, ANN(1) resulted to be significantly better in the classification of reassuring lesions than ANN(2). This study suggests that multispectral image analysis and artificial neural networks could be used to support primary care physicians or general practitioners in identifying pigmented skin lesions that require further investigations.