Automated dermoscopy image analysis

Can automated dermoscopy image analysis instruments provide added benefit for the dermatologist? A study comparing the results of three systems BJD Nov 2007 A. Perrinaud, O. Gaide**Department of Dermatology, University Hospital Geneva, Geneva, Switzerland, L.E. French††Department of Dermatology, University Hospital Zurich, Gloriastrasse 31, CH-8091 Zurich, Switzerland, J.-H. Saurat**Department of Dermatology, University Hospital Geneva, Geneva, Switzerland, A.A. Marghoob‡‡Dermatology Service, Memorial Sloan-Kettering Cancer Center, New York, NY, U.S.A. and R.P. Braun*†Department of Dermatology, University Hospital Trousseau, Tours, France*Department of Dermatology, University Hospital Geneva, Geneva, SwitzerlandDepartment of Dermatology, University Hospital Trousseau, Tours, France
*Department of Dermatology, University Hospital Geneva, Geneva, Switzerland
†Department of Dermatology, University Hospital Zurich, Gloriastrasse 31, CH-8091 Zurich, Switzerland
‡Dermatology Service, Memorial Sloan-Kettering Cancer Center, New York, NY, U.S.A.
Ralph P. Braun.
E-mail: ralph.braun@usz.ch


Summary

Background Instruments designed to provide computer program-driven diagnosis of dermoscopic images of lesions are now commercially available. Multiple publications tout the improved diagnostic accuracy of these instruments compared with that of clinicians.

Objectives Our aim was to evaluate the actual usefulness of these instruments for dermatologists practising in a pigmented lesion clinic.

Methods Over a 4-month period we subjected lesions, which were being evaluated in one of our clinics, to automated computer diagnosis performed by three commercially available instruments. We intentionally included three groups of lesions: group 1 lesions were suspicious melanocytic lesions that were scheduled to be excised; group 2 lesions were nonmelanocytic lesions; group 3 lesions were clinically obvious melanomas. The automated diagnoses provided by the instruments were compared with the dermoscopy diagnosis of experienced physicians and with histopathology.

Results We included a total of 107 lesions. One imaging system’s computer algorithm was unable to analyse one third of the lesions. All three instruments’ computer algorithms were able to identify the clinically obvious melanomas (group 3) correctly. However, all three systems tended to overdiagnose by incorrectly classifying most seborrhoeic keratoses (group 2) as potential malignant lesions. Concerning the suspect melanocytic lesions (group 1), which are precisely the lesions for which a dermatologist would welcome a second opinion, we found significant variability in the diagnostic accuracy of the instruments tested. However, all three systems providing computer-assisted diagnosis had a tendency to overdiagnose benign melanocytic lesions as potential melanomas.

Conclusions Although the image analysis systems tested by us correctly identified the clinically obvious melanomas, they were not able to discriminate between most dysplastic naevi and early malignant melanoma. Thus, for the moment these computer-assisted diagnostic imaging machines provide little to no added benefit for the experienced dermatologist/dermoscopist.