Díaz de María, FernandoLópez Labraca, Javier2018-04-162018-04-1620142014-07-11https://hdl.handle.net/10016/26651Computer Aided Diagnosis (CAD) systems for melanoma detection aim to mirror the expert dermatologist decision when watching a dermoscopic or clinical image. Computer Vision techniques, which can be based on expert knowledge or not, are used to characterize the lesion image. This information is delivered to a machine learning algorithm, which gives a diagnosis suggestion as an output. This research is included into this field, and addresses the objective of implementing a complete CAD system using ‘state of the art’ descriptors and dermoscopy images as input. Some of them are based on expert knowledge and others are typical in a wide variety of problems. Images are initially transformed into oRGB, a perceptual color space, looking for both enhancing the information that images provide and giving human perception to machine algorithms. Feature selection is also performed to find features that really contribute to discriminate between benign and malignant pigmented skin lesions (PSL). The problem of robust model fitting versus statistically significant system evaluation is critical when working with small datasets, which is indeed the case. This topic is not generally considered in works related to PSLs. Consequently, a method that optimizes the compromise between these two goals is proposed, giving non-overfitted models and statistically significant measures of performance. In this manner, different systems can be compared in a fairer way. A database which enjoys wide international acceptance among dermatologists is used for the experiments.application/pdfengAtribución-NoComercial-SinDerivadas 3.0 EspañaBiomarkersComputer-aided diagnosisImage processingMelanoma analysisDermatologistsbachelor thesisBiología y Biomedicinaopen access