In this chapter, the authors propose an approach for building a model characterizing malignant melanomas. A common way to build a domain model is using an inductive learning method. Such resulting model is a generalization of the known examples. However, in some domains where there is not a clear difference among the classes, the inductive model could be too general. The approach taken in this chapter consists of using lazy learning methods for building what the authors call a lazy domain theory. The main difference between both inductive and lazy theories is that the former is complete whereas the latter is not. This means that the lazy domain theory may not cover all the space of known examples. The authors’ experiments have shown that, despite of this, the lazy domain theory has better performance than the inductive theory.
TopIntroduction
Malignant melanoma (MM) is the most dangerous type of skin cancer. The MM is the second most frequent kind of cancer among people between 15 and 34 years old. In the last thirty years the incidence of MM has been increased more rapidly than other kinds of cancer. Many studies show that an early detection of MM increases the survival rate since when tumors are thin the lesion can be excised and the survival is around the 95% after 5 years. However, when the tumor has spreading to the nodes the risk of metastases increases and, thus the survival rate decreases. The early diagnosis of melanoma is a difficult task that dermatologist face every day. When a lesion is suspicious of being a melanoma it is removed and the final diagnosis is performed based on histopathology criteria.
The clinical diagnosis of MM is based on the ABCD rule that takes into account the Asymmetry, Border irregularity, Color and Diameter of the lesion. Although the ABCD rule has been proved to be effective for an early diagnosis, there are necessary more accurate methods to correctly diagnose lesions that do not present clear malignant characteristics. It is important that a dermatologist can detect suspicious skin lesions during a clinical session, therefore it should be very useful to have a clear and easy characterization of MM in early stages. Dermoscopy is a non-invasive technique introduced by dermatologists two decades ago. This technique provides a more accurate evaluation of skin lesions, and can therefore, avoid the excision of lesions that are benign. Therefore, dermatologists need to achieve a good dermatoscopic classification of lesions prior to extraction (Puig et al., 2007). Hofmann-Wellenhof et al. (2002) suggested a classification of benign melanocytic lesions. Argenziano et al. (2007) hypothesized that dermoscopic classification may be better than the classical clinico pathological classification of benign melanocytic lesions (nevi). Dermoscopy improves accuracy for the diagnosis of melanoma in nearly 25%. However, some benign lesions may mimic melanoma and some melanomas may be similar to benign lesions, consequently many unnecessary extractions are produced. It is assessed that one of 30 lesions excised by non-expert dermatologists only one of them is MM. When dermatologists have high expertise, the ratio decreases to one MM for each 4 excisions. For this reason, the benign/malignant ratio of excised lesions is 1 malignant for 30 benign for non-expert dermatologists, whereas this ratio is 1 MM to 4 benignant when the dermatologist has high expertise. The reflectance confocal microscopy is a new non-invasive diagnostic technique that allows the visualization of skin cells in vivo. This technique also increase the accuracy of the experts diagnosis but even in the hands of experts and in combination with dermoscopy information, accuracy never reaches 100%.
Thus, we are especially interested on characterizing skin lesions in the frontier of both malignant and benignant lesions. In our experiments we used descriptions of skins lesions that have already been excised, i.e., they are lesions that dermatologists considered that could be malignant melanomas. However some of them, after a histopathology analysis resulted to be benignant. This means that they provide a good set of suspicious lesions from which to generate a domain model able to discriminate between both malignant and benignant lesions with similar characteristics. We propose to take descriptions of known skin lesions and to use a lazy learning method to obtain a domain theory. Skin lesions are described using two sets of features, dermatoscopic and confocal, and our goal is to find a subset of features characterizing malignant lesions.
What we propose is to use lazy learning methods for building a domain theory useful for the classification of skin lesions. We experimented with two lazy learning methods: k nearest neighbor and Lazy Induction of Descriptions (LID). The k nearest neighbor (k-NN in short) method is based on the idea that similar objects have similar classification. Once an unseen object has been classified, we propose to build an explanation of the classification based on the anti-unification concept, i.e., the explanation is formed by the attributes shared by all the objects assessed as the most similar to the unseen object. The LID method is a lazy learning method useful for classification tasks. In addition of classifying a new domain object, LID also gives a symbolic description that can be interpreted in several ways. One of these interpretations is as a partial description of a class.