Inductive Machine Learning for Classifying Breast Lesions from Imbalanced MRI-B Data

This paper presents a hybrid machine-learning approach for modelling the diagnostic process of breast cancer by analyzing data derived from medical observations of magnetic mammograms. Specifically, magnetic mammography data MRI-data) collected in collaboration with medical experts, are utilized. The proposed methodology combines different stages of data pre-processing (value grouping of variables, filling-in missing values, etc.) in combination to expert knowledge and guidance. The range of the involved diagnostic variables is identified with the aid of descriptive statistics (frequencies), as well as the number of training instances for all the decision classes. Fifteen decision variables are used, both nominal and numerical, corresponding to imaging data and clinical observations. Two different decision problems are defined and solved, (i) a 2-class problem corresponding to the discrimination between healthy and cancerous situations, and (ii) a 6-class problem corresponding to the separation between different specialized cancerous diagnoses. A new data balancing technique is then applied, adapted to the special requirements of the specific application domain, aiming at the best possible distribution of training instances based on entropy information criteria. In addition, boosting techniques are also applied in the phase of rule induction. Feature selection techniques can also assist the whole process by reducing possible noise existing in the data through removal of confusing variables. Diagnostic rules are finally induced either in the form of decision trees or in the form of rule sets, which obtain higher classification accuracy than standard ML approaches, while comprehensibility of the outcome remains high. Specifically, results are compared with the use of known standard metrics (k-fold cross-validation, leave-one-out and use-test-set approaches) as well as by related medical experts. The classification performance is very satisfactory (exceeds 95%) and the comments of the experts are encouraging. Indicative rules are presented and discussed in the paper, while future directions are pointed out.