The iDScore method

Dermoscopy is a non-invasive method which proved to enhance skin lesions diagnosis. During the last decades, many dermoscopic checklists, algorithms, rules, methods, etc. have been proposed in order to improve the accuracy of early dermoscopic diagnosis of melanoma and to avoid inappropriate removal. However, none of the available dermoscopic algorithms appear able to reach very high sensitivity and specificity levels in the differentiation of atypical/dysplastic nevi (AN) with worrisome features form early melanoma (EM). Thus, the non-invasive diagnosis of these clinically and dermosocpically difficult melanocytic skin lesions (MSLs) rests a challenge in daily practice. Actually, a series of clinical and personal data of the lesion/patient orient our managment decision in routine visists, but no algorithms to date were able to statistically evaluate the impact of these data on the final decision of clinicians (1).

For this purpose, we developed and validated a new method able to simultaneously combine dermoscopy data with personal data of the patient. by means of a scoring classifier statistical method (2): the integrated Dermoscopy Score model – the iDScore. This integrated score algorithm can calculate a patient’s risk score for melanoma, as it is very effective in selecting the significant parameters for discriminating two clinical conditions, deleting all redundant variables. When examining a patient with suspected MSL, the dermatologist could rapidly reach a non invasive differential diagnosis between AN and EA with a high confidence level: indeed, the individual diagnostic abilities of the dermatologists will be enhanced when using the proposed algorithm, offering performance far superior to their standard pattern analysis.

This method was succesfully applied to regressing nevi and melanoma with regression reaching high diagnostic accuracy in 2015 (3), then to the atypical nevi from early melanomas of all body sites (excluding face, palms and plants) in 2016 (4).

Few relevant clinical-personal data were selected and coded accordign to a specifica risk progression scale, that are: age (4 age range groups), lesions diameter (3 diamter groups) and body site (14 sites grouped according to a specifically developed sun-exposure based classification). The risk scale range was 0-16, so distributed: S=0-3: No risk 5; S=4-5: very low risk; S=6-7: low risk; S=8-10: moderate risk; S=11-13: high risk; S=14-16: very high risk. Corresponding management suggestion are proposed to these 6 risk ranges: no follow-up/long follow-up/medium follow-up/short follow-up/very short follow-up/immediate excision.

PPRC EADV project 2017-2019

The project workflow included the following steps: construction of a web platform dedicated to teledermoscopy (TWP); collection of a large database of 981 MSL with clinical and dermoscopical features of malignancy, (i.e., 664 AN and 317 EM) including: 981 dermoscopic standardized images, 390 clinical images, histological diagnosis, clinical data (i.e., lesion max diameter, body area localtion) and personal data (age and gender of the patient) i.e., http://eadv-idscore.dbm.unisi.it; definition of dermoscopic criteria of malignancy able to separate AN from EM on standard dermoscopy analysis; statistical analysis; elaboration of the integrated score model and successive testing set over up to 100 participants dermatologists of difference experience levels; 6-months testing session period on the TWP: http://idscore.dbm.unisi.it; comparison diagnostic performances obtained by each participant with standard dermoscopic pattern analysis and with score algorithm-aided diagnosis.

The iDSCore checklist 2016 was so validated in a large multicentric dataset of lesion in a teledermatology setting, and compared with intuitive diagnosis based on pattern analysis, the ABCD rule and the 7-point checklist: the diagnostic accuracy reached was higher than the other methods for all tested lesions (~1000)(5). Dermatologists of all experience in dermoscopy degree showed to benefit from the application of the iDScore algorithm, iuncluding: skill level I (<1 year-experience), skill level II (1-4 year-experience), skill level III (6-8 year-experience), skill level IV (>8 year-experience).

The evolution to "Integrated Dermatological Scoring system classifiers"

The technical advances in non-invasive imaging in the last decades has been outstanding.

We have the possibility to “explore” MSLs, in vivo, navigating at 400x magnification with high resolution dermoscopes, reaching cellular resolution with reflectance confocal laser microscopes (RCM) or to visualizing the whole lesion structure with skin Optical coherence tomography (OCT). On the oher hand, all these techniques still need the comparison with histopathologic examination, that rests the gold standard. More recently, a new techinique that combines the two, named line field optical coherence tomography (LC-OCT) was developed, and could possibly considered as an in vivo histology. (6,7)

Beside this, the artificial intelligence (AI) studies are more and more interested in supporting dermatologists in the differential diagnosis of AN and EM, in particular deep convolutional neural network specifically trained.(8)

We first aimed to realize a web platform that allows dermatologists to compare multiple images of MSLs appearing on of the whole body surface obtained with different non-invasive devices, namley:

• Standard dermoscopy

• High resolution dermoscopy 400x

• Reflectance confocal microsocpy (RCM)

• Line field optical coherence tomography (LC-OCT)

A second aim is to combine multiple clinical, personal and imaging data obtained with these devices in order to select those relevant in the distinguish EM from AN.

A third aim is to develop and validate new iDScore algorithms specifically designed with using one or more non-invasive tool.

A fourth aim is to evaluate a possible integration of these selected significant data into a deep convolution neural network, in order to create an hybrid AI model with increased diagnostic accuracy and those responses can be, at least partially, interpreted by the human mind / the dermatologist.

1. Tognetti L, Cinotti E, Moscarella E, Farnetani F, Malvehy J, Lallas A, Pellacani G, Argenziano G, Cevenini G, Rubegni P. Impact of clinical and personal data in the dermoscopic differentiation between early melanoma and atypical nevi. Dermatol Pract Concept. 2018 Oct 31;8(4):324-327.
   https://www.ncbi.nlm.nih.gov/pubmed/30479866

2. Cevenini G, Furini S, Barbini P, Tognetti L, Rubegni P. Scoring systems in dermatology. 2016 IEEE International Symposium on Medical Measurements and Applications (MeMeA), Benevento, 2016, pp. 1-6.
   https://ieeexplore.ieee.org/document/7533793?isnumber=7533693&arnumber=7533793&tag=1

3. Rubegni P, Tognetti L, Argenziano G, Nami N, Brancaccio G, Cinotti E, Miracco C, Fimiani M1, Cevenini G. A risk scoring system for the differentiation between melanoma with regression and regressing nevi. J Dermatol Sci. 2016 Aug;83(2):138-44.
   https://www.ncbi.nlm.nih.gov/pubmed/27157925

4. Tognetti L, Cevenini G, Moscarella E, Cinotti E, Farnetani F, Mahlvey J, Perrot JL, Longo C, Pellacani G, Argenziano G, Fimiani M, Rubegni P. An integrated clinical-dermoscopic risk scoring system for the differentiation between early melanoma and atypical nevi: the iDScore. J Eur Acad Dermatol Venereol. 2018 Dec;32(12):2162-2170.
   https://www.ncbi.nlm.nih.gov/pubmed/29888421

5. Tognetti L, Cevenini G, Moscarella E, Cinotti E, Farnetani F, Lallas A, Tiodorovic D, Carrera C, Puig S, Perrot JL, Longo C, Argenziano G, Pellacani G, Smargiassi E, Cataldo G, Cartocci A, Balistreri A, Rubegni P. Validation of an integrated dermoscopic scoring method in an European teledermoscopy web platform: the iDScore project for early detection of melanoma. J Eur Acad Dermatol Venereol 2019
   https://www.ncbi.nlm.nih.gov/pubmed/31465600

6. Dubois A, Levecq O, Azimani H, Davis H, Ogien J, Siret D, Barut A. Line-field confocal time-domain optical coherence tomography with dynamic focusing," Opt. Express 26, 33534-33542 (2018)
   https://www.osapublishing.org/oe/abstract.cfm?uri=oe-26-26-33534

7. Dubois A, Levecq O, Azimani H, Davis H, Siret D, Barut A, Suppa M, del Marmol V, Malvehy J, Cinotti E, Rubegni P, Perrot JL. Line-field confocal optical coherence tomography for high-resolution noninvasive imaging of skin tumors," J. Biomed. Opt. 23(10) 106007 (23 October 2018)
   https://www.ncbi.nlm.nih.gov/pubmed/30353716

8. Fink C, Blum A, Buhl T, et al. Diagnostic performance of a deep learning convolutional neural network in the differentiation of combined nevi and melanomas. J Eur Acad Dermatol Venereol. 2019 Dec 19.
   https://www.ncbi.nlm.nih.gov/pubmed/31856342