Applying Association Rule Mining to Semantic Data in the Lung Image Database Consortium

The detection and diagnosis of lung cancer has been shown to dramatically increase the survival rate of lung cancer patients. Computer Aided Diagnosis (CAD) methods are being developed to help improve the ability of clinical radiologists to detect and diagnose malignant lung nodules. While many research studies use low-level image features to predict malignancy, only few CAD systems look into integrating semantic data (such as spiculation, subtlety and margin characteristics) in these systems. The availability of these semantic characteristics in the NIH/NCI Lung Nodule Database Consortium (LIDC) creates new opportunities to explore the relationships among these semantic characteristics in the context of lung nodule diagnostic interpretation. We propose the use of Association Rule Mining (ARM) to quantify these relationships and introduce new evaluation metrics that relate to the importance of individual characteristics rather than to a set of rules as a whole. Our preliminary results show that, although there is less evidence that malignancy can be predicted based on the other semantic characteristics, there is strong support and confidence for the existence of certain combinations of characteristics (including malignancy) that could be used to identify groups of nodules that are described in a similar fashion by radiologists.