Utilizzo del Deep Learning nella Analisi di Immagini Dermoscopiche

Abstract: Countless different imaging acquisition techniques are employed by medical practitioners as a tool to perform diagnosis, ranging from microscopy to Magnetic Resonance Imaging (MRI). This common practice produced a great opportunity for computer vision algorithms to find ways to perform automated analysis on medical images. In particular, after the groundbreaking success of AlexNet in 2012, deep learning has become a vital element in medical imaging research. Above all, Convolutional Neural Networks (CNNs) have been successfully adopted to perform a great variety of tasks such as image segmentation, classification, detection, and generation. This thesis is a collection of deep learning applications for dermoscopic images analysis. Dermoscopy is a form of in-vivo skin surface microscopy performed using high quality magnifying lenses and a powerful light source to mitigate the surface reflection of the skin, to enhance the visibility of the pigmentation of the lesion. However, to fully make use of this non-invasive imaging approach, a thorough image analysis must be performed by expert clinicians, and therefore many efforts have been given in recent years towards the creation of tools to assist physicians in the analysis of dermoscopic images. We start by approaching lesion segmentation, by means of a novel data augmentation technique and a diverse ensemble strategy. The final goal of dermoscopic images analysis is skin lesion classification, for which we develop an approach that achieved the third best result in the 2019 ISIC global challenge. Moreover, we address one of the main drawbacks of deep learning algorithms, their low interpretability, by using content-based image retrieval to assist the diagnosis process of both expert and novice practitioners and, finally, by trying to determine which characteristics are taken into account by autonomous classification algorithms.

 not available