Classification of melanomas in situ using knowledge discovery with explained case-based reasoning.

OBJECTIVE: Early diagnosis of melanoma is based on the ABCD rule which considers asymmetry, border irregularity, color variegation, and a diameter larger than 5mm as the characteristic features of melanomas. When a skin lesion presents these features it is excised as prevention. Using a non-invasive technique called dermoscopy, dermatologists can give a more accurate evaluation of skin lesions, and can therefore avoid the excision of lesions that are benign. However, dermatologists need to achieve a good dermatoscopic classification of lesions prior to extraction. In this paper we propose a procedure called LazyCL to support dermatologists in assessing the classification of skin lesions. Our goal is to use LazyCL for generating a domain theory to classify melanomas in situ.

METHODS: To generate a domain theory, the LazyCL procedure uses a combination of two artificial intelligence techniques: case-based reasoning and clustering. First LazyCL randomly creates clusters and then uses a lazy learning method called lazy induction of descriptions (LID) with leave-one-out on them. By means of LID, LazyCL collects explanations of why the cases in the database should belong to a class. Then the analysis of relationships among explanations produces an understandable clustering of the dataset. After a process of elimination of redundancies and merging of clusters, the set of explanations is reduced to a subset of it describing classes that are "almost" discriminant. The remaining explanations form a preliminary domain theory that is the basis on which experts can perform knowledge discovery.

RESULTS: We performed two kinds of experiments. First ones consisted on using LazyCL on a database containing the description of 76 melanomas. The domain theory obtained from these experiments was compared on previous experiments performed using a different clustering method called self-organizing maps (SOM). Results of both methods, LazyCL and SOM, were similar. The second kind of experiments consisted on using LazyCL on well known domains coming from the machine learning repository of the Irvine University. Thus, since these domains have known solution classes, we can prove that the clusters build by LazyCL are correct.

CONCLUSIONS: We can conclude that LazyCL that uses explained case-based reasoning for knowledge discovery is feasible for constructing a domain theory. On one hand, experiments on the melanoma database show that the domain theory build by LazyCL is easy to understand. Explanations provided by LID are easily understood by domain experts since these descriptions involve the same attributes than they used to represent domain objects. On the other hand, experiments on standard machine learning data sets show that LazyCL is a good method of clustering since all clusters produced are correct.