Contributions to melanoma computer aided diagnosis systems using dermoscopic images

Abstract

Melanoma is the most dangerous form of skin cancer and it has a very low survival rate when treated in advanced stages. However, the prognosis is much more hopeful if it is detected in early stages. Dermoscopy is a well known skin lesion visualization technique which reveals many structural details that are hidden to the naked eye without the need of taking a biopsy. There is an increasing interest in the development of computer-aided systems to assist the clinicians in their diagnosis of this form of cancer using dermoscopy. However, despite of the many research efforts devoted to this matter, these systems have not yet been incorporated into the daily clinical practice. One of the underlying reasons for this lack of success is that most of the systems simply provide a parallel diagnosis to the clinicians, which actually does not help them in practice. Although very powerful in terms of classification accuracy, most of these systems lack a clear understanding of the main underlying factors and lesion properties that support their final decision. This causes the clinicians not to trust the output of these automatic systems. In this Ph.D. Thesis, we have proposed two Computer-aided Diagnosis systems for early melanoma detection. Our first approach aims to go beyond a simple diagnosis suggestion, to provide enriched information about its automatic diagnosis. We have designed a novel set of clinical-based features, which are based on both dermoscopic structures and diagnosis criteria from dermoscopic algorithms. The medical basis allows these features to show both competitive performance in melanoma detection and at the same time good interpretability for dermatologists. Moreover, we have focused on obtaining quantitative measures of clinical criteria, which provides clinicians with additional information to perform lesion diagnosis. In order to present this information to the clinician, we have designed an automatic tool. It shows the segmentation into the dermoscopic structures that are present in the lesion, together with the weight of each structure-based expert on the final decision and the features having stronger influence in the decision of each expert. The performance of the first system has been assessed in two scenarios. On the one hand, we use a dataset containing difficult cases for dermatologists, in order to emulate the behavior of our system in the clinical practice of an expert clinician. In this case, it has been shown that there is a boost in performance when combining our system with the diagnosis of an expert dermatologist, which demonstrates the usefulness of the system as a support element to the decisions of dermatologists. On the other hand, our approach improves the performance of quite a few competent methods from the state-of-the-art. Although it provides results just close to those of CNN-based methods, it offers an understandable interface for clinicians, so they can obtain complementary information and know the main reasons that led the system to make its diagnostic decision. Our second proposal is based on CNNs, since these algorithms have shown to be more powerful than classical approaches in many Computer Vision scenarios, including skin lesions diagnosis. Although their interpretability is very limited in comparison with a classical system, some efforts can be made in this direction so the systems become more practical and not only better in terms of performance. According to this idea, we have adapted the architecture of a typical image classification network to incorporate dermoscopic structures explicitly, in order to guarantee their contribution to the diagnosis suggestion. To this purpose, we have used a structure segmentation network to obtain the dermoscopic structure maps, which are then fed to a diagnosis network. The diagnosis network considers both the original dermoscopic image and the structure maps, and yields a diagnosis proposal. The proposed segmentation network is able to learn from both strong and weak labels, using a hybrid loss function. In our experiments, we have checked the segmentation performance using different proportions of strong and weak labels. The results show that, once a minimum number of strongly-labeled examples are gathered, the rest of the effort can be allocated to obtaining weakly-labeled samples. Weak labels are cheap to obtain for a dermatologist, while strong labels are much more costly. This implies that the cost of building a database for dermoscopic structure segmentation should be reduced. With a correct balance between weak and strong labels, the performance would not be significantly affected. Regarding the diagnosis network, it is formed by a main network, which receives the dermoscopic image of the lesion, and an auxiliary network that gets the structure segmentations and builds feature representations to incorporate this information at several selected locations of the main network. Our experimental results show that the proposed method provides better average results than a selected set of state-of-the-art methods. Furthermore, it also outperforms two relevant ensemble-based methods from the ISIC challenge (an important international contest on melanoma diagnosis using dermoscopic images), which suggests that the diagnosis information embedded into the dermoscopic structures is not properly inferred by the learned features of general purpose architectures, even for very complex models. Moreover, the proposed system improves the results of other state-ofthe- art methods in terms of specificity at 95% sensitivity, which is a performance measure well adapted to the practical daily problem of filtering benign lesions for having the doctors focused on suspicious lesions.