Automatic colour correction of digital skin images in teledermatology

Abstract

We have developed an algorithm for automatic colour correction of skin images obtained in imperfect conditions. The algorithm has been incorporated into a computer program, TransImage. In a set of 31 digital skin images (including cases of dermatitis, eczema, mycotic lesions and skin cancer), 19 images underwent colour correction with the computer program, while 12 remained uncorrected. The images were presented to three experienced dermatologists who reviewed them without knowing if they had been corrected or not. The mean colour accuracy score (on a scale from 1=very poor to 5=excellent) for the corrected skin images was 4.1, while the mean score for uncorrected images was 3.7; the difference was significant (P<0.05). Using the 19 corrected images, the dermatologists made 11 inaccurate diagnoses out of 57 (a mean diagnostic accuracy of 81%). Using the 12 uncorrected images, three dermatologists made 16 inaccurate diagnoses out of 36 (i.e. a mean diagnostic accuracy of 56%); the difference was significant (P<0.05). Colour correction increased the diagnostic accuracy of teledermatology.