Apprentissage profond pour l’aide au diagnostic du mélanome à partir d’exemple

Arthur Cartel Foahom Gouabou (1, 2, 3) , Frédéric Heim (4, 2) , Jean-Luc Damoiseaux (1, 5)

1 LIS - Laboratoire d'Informatique et des Systèmes (LIS) (Marseille, Toulon)
2 Université de Haute-Alsace (UHA) - Université de Haute-Alsace (UHA) Mulhouse - Colmar
3 I&M - Images et Modèles
4 LPMT - Laboratoire de Physique et Mécanique Textiles
5 AMU - Aix Marseille Université

One study reveals that 15404 new cases of cutaneous melanoma have been estimated in France in 2017. The 5-year survival rate of a person with advanced melanoma is much lower than 20%, which raises the need for diagnose it at an early stage. The purpose of this work is to build a supervised computer-aided diagnosis system for melanoma. The database used for the implementation includes 1356 images divided into 9 classes. Two approaches have been implemented : classical approach and deep learning approach. The classical approach combines two support vector machine classifiers (SVM) trained on features extracted from three extractors. This approach yielded an area under the receptor curve (AUC) of 0.88, a sensitivity (SE) of 89% and a specificity (SPEC) of 77%. The deep learning approach uses features extracted from two pre-trained models VGG16 and resnet50 to train two linear SVM. The scores from these two classifiers are combined using a logistic regression algorithm to obtain the classification. This approach yielded an AED-CCR of 0.88, SE of 78% and SPEC of 83%.

Mots clés: Melanoma detection; Support vecteur machines; Deep learning; Features extraction

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]; Intelligence artificielle [cs.AI]; Imagerie médicale; Traitement des images [eess.IV]; Réseau de neurones [cs.NE]

Arthur Cartel Foahom Gouabou, Frédéric Heim, Jean-Luc Damoiseaux. Apprentissage profond pour l’aide au diagnostic du mélanome à partir d’exemple. [Rapport de recherche] Université de Haute-Alsace. 2019. ⟨hal-02515203⟩