Soft computing and artificial intelligence for interpretable analysis of dermoscopic images

Resumen

In this thesis, we apply deep learning techniques to dermoscopic images of skin lesions. Artificial intelligence, and more specifically deep learning, have had a major impact on the computer vision community in many fields. In dermatology, artificial intelligence has achieved dermatologist-level accuracy for skin cancer classification. Although melanoma remains an incurable disease, the survival rate and the efficacy of treatment increase greatly if detected early. This work aims to provide automatic tools to help physicians assess the malignancy of a skin lesion, including: 1. An encoder-decoder model for the detection and posterior restoration of hair’s pixels from dermoscopic images is designed. The existence of hairs in these images may occlude relevant patterns and hinder the lesion assessment. A statistical test states the superiority of our model in eight of the nine similarity measures when compared to state-of-the-art methods. In addition to good quantitative results on images with simulated hair, excellent visual results have been obtained on images with real hair. 2. A deep learning approach to classify the skin lesions as “fully asymmetric”, “symmetric with respect to one axis”, or “symmetric with respect to two axes” is introduced, due to the clinical significance of the asymmetry in assessing the malignancy of lesions. Compared to traditional methods, our proposed method largely outperforms them, even when the task is simplified to a binary problem, and benefits from not requiring the lesion segmentation. However, it has not been able to generalize well when presented with data from another dermatological database. 3. Finally, two multitask learning systems have been built to provide more context for the specialist to make a decision and rely on our system. The first one focuses on simultaneously segmenting the skin lesion and the hairs present in the image, as well as performing the inpainting of these hair’s regions. The second model combines the tasks of lesion segmentation; and its classification according to their symmetry; their diagnosis; and their malignancy. In both cases, we have also studied how different combinations of these tasks influence each other. We found that in some cases the tasks benefited from the multitask environment but in others such as inpainting or symmetry classification, the multitask performance results are lower than their solo performance. in particular, closely related tasks (e.g. diagnosis and malignancy classification) tend to improve their performance when learned simultaneously. Finally, we found that adding related tasks acts as a learning regularization, improving in some cases the speed of convergence. Also, we introduce two datasets. The first one consists of images with simulated hair, while the second one, contains the annotations of three experts regarding the symmetry of lesion. In both cases, we use images extracted from public datasets.