Partial order label decomposition approaches for melanoma diagnosis

• Published in: Applied soft computing 2018-03, Vol.64, p.341-355
• Main Authors: Sánchez-Monedero, Javier, Pérez-Ortiz, María, Sáez, Aurora, Gutiérrez, Pedro Antonio, Hervás-Martínez, César
• Format: Article
• Language: English
• Subjects: Computer vision; Machine learning; Melanoma; Ordinal classification; Partial order; Skin cancer
• ISSN: 1568-4946; 1872-9681
• DOI: 10.1016/j.asoc.2017.11.042

•We achieve both melanoma detection and fine grain melanoma stage classification.•Based on the problem definition, we propose partial ordered models and oversampling methods.•The proposed topology-aware methods improved the accuracy and reduced the magnitude of the errors.•Analysis of models brings insights of the contribution of each feature to the classification task. Melanoma is a type of cancer that develops from the pigment-containing cells known as melanocytes. Usually occurring on the skin, early detection and diagnosis is strongly related to survival rates. Melanoma recognition is a challenging task that nowadays is performed by well trained dermatologists who may produce varying diagnosis due to the task complexity. This motivates the development of automated diagnosis tools, in spite of the inherent difficulties (intra-class variation, visual similarity between melanoma and non-melanoma lesions, among others). In the present work, we propose a system combining image analysis and machine learning to detect melanoma presence and severity. The severity is assessed in terms of melanoma thickness, which is measured by the Breslow index. Previous works mainly focus on the binary problem of detecting the presence of the melanoma. However, the system proposed in this paper goes a step further by also considering the stage of the lesion in the classification task. To do so, we extract 100 features that consider the shape, colour, pigment network and texture of the benign and malignant lesions. The problem is tackled as a five-class classification problem, where the first class represents benign lesions, and the remaining four classes represent the different stages of the melanoma (via the Breslow index). Based on the problem definition, we identify the learning setting as a partial order problem, in which the patterns belonging to the different melanoma stages present an order relationship, but where there is no order arrangement with respect to the benign lesions. Under this assumption about the class topology, we design several proposals to exploit this structure and improve data preprocessing. In this sense, we experimentally demonstrate that those proposals exploiting the partial order assumption achieve better performance than 12 baseline nominal and ordinal classifiers (including a deep learning model) which do not consider this partial order. To deal with class imbalance, we additionally propose specific over-sampling techniques that consider the s