6533b7d9fe1ef96bd126be65

RESEARCH PRODUCT

Classification par méthodes d’apprentissage supervisé et faiblement superviséd’images multimodales pour l’aide au diagnostic du lentigo malin en dermatologie

subject

description

Carried out in collaboration with the Saint-Étienne University Hospital, this work provides additional information to help the skin diagnosis by providing new decision methods on Lentigo Maligna and Lentigo Maligna Melanoma pathologies. To this end, the modalities regularly used in clinical conditions are made available to this work and are orchestrated within a multimodal process. Among image modalities, may be mentioned the clinical photography, the dermatoscopy, and the confocal reflectance microscopy. Initially, the first steps of this manuscript focus on reflectance confocal microscopy as the work in computer diagnostic assistance is relatively underdeveloped, in particular on the detection of pathologies of Lentigo Maligna and Lentigo Maligna Melanoma. To this end, several automatic learning methods are described to allow the separation of this modality at the level of the image data according to the healthy, benign, and malignant annotations. Then, the most image-relevant methods are combined with supervised and weakly supervised prediction methods to enable decision-making at the lesion level. Secondly, the proposed multimodal process resumes the clinical path, starting with the most affordable and accessible modality, namely clinical photography, dermatoscopy, and finally reflectance confocal microscopy, respectively. According to this principle, two schemes are proposed in this manuscript, one considered as without memory, i.e. requiring only the data made available by the current modality, and the other with memory, i.e. by an accumulation of information acquired previously.