Dynamic clustering detection through multi-valued descriptors of dermoscopic images.
Stat Med
; 30(20): 2536-50, 2011 Sep 10.
Article
en En
| MEDLINE
| ID: mdl-21717490
This paper introduces a dynamic clustering methodology based on multi-valued descriptors of dermoscopic images. The main idea is to support medical diagnosis to decide if pigmented skin lesions belonging to an uncertain set are nearer to malignant melanoma or to benign nevi. Melanoma is the most deadly skin cancer, and early diagnosis is a current challenge for clinicians. Most data analysis algorithms for skin lesions discrimination focus on segmentation and extraction of features of categorical or numerical type. As an alternative approach, this paper introduces two new concepts: first, it considers multi-valued data that scalar variables not only describe but also intervals or histogram variables; second, it introduces a dynamic clustering method based on Wasserstein distance to compare multi-valued data. The overall strategy of analysis can be summarized into the following steps: first, a segmentation of dermoscopic images allows to identify a set of multi-valued descriptors; second, we performed a discriminant analysis on a set of images where there is an a priori classification so that it is possible to detect which features discriminate the benign and malignant lesions; and third, we performed the proposed dynamic clustering method on the uncertain cases, which need to be associated to one of the two previously mentioned groups. Results based on clinical data show that the grading of specific descriptors associated to dermoscopic characteristics provides a novel way to characterize uncertain lesions that can help the dermatologist's diagnosis.
Palabras clave
Texto completo:
1
Colección:
01-internacional
Base de datos:
MEDLINE
Asunto principal:
Neoplasias Cutáneas
/
Algoritmos
/
Dermoscopía
/
Melanoma
Tipo de estudio:
Diagnostic_studies
/
Prognostic_studies
/
Screening_studies
Límite:
Humans
Idioma:
En
Revista:
Stat Med
Año:
2011
Tipo del documento:
Article
País de afiliación:
Italia
Pais de publicación:
Reino Unido