Comparative Analysis of Distributed, Default, IC, and Fuzzy ARTMAP Neural Networks for Classification of Malignant and Benign Lesions

Only one third of all breast cancer biopsies made today confirm the disease, which make these procedures inefficient and expensive. We address the problem by exploring and comparing characteristics of four neural networks used as predictors: fuzzy, distributed, default, and ic ARTMAP, all based on the adaptive resonance theory. The networks were trained using a dataset that contains a combination of 39 mammographic, sonographic, and other descriptors, which is novel for the field. We compared the model performances by using ROC analysis and metrics derived from it, such as max accuracy, full and partial area under the convex hull, and specificity at 98% sensitivity. Our findings show that the four models outperform the most popular MLP neural networks given that they are setup properly and used with appropriate selection of data variables. We also find that two of the models, distributed and ic, are too conservative in their predictions and do not provide sufficient sensitivity and specificity, but the default ARTMAP shows very good characteristics. It outperforms not only its counterparts, but also all other models used with the same data, even some radiologist practices. To the best of our knowledge, the ARTMAP neural networks have not been studied for the purpose of the task until now.