Abstract: This study explores and compares predictive abilities of six types of neural networks used as tools for computer-aided breast cancer diagnosis, namely, multilayer perceptron, cascade-correlation neural network, and four ART-based neural networks. Our experimental dataset consists of 803 patterns of 39 BI-RADS, mammographic, sonographic, and other descriptors. Using such a combination of features is not traditional in the field and we find it is better than traditional ones. The study also focuses on exploring how various feature selection techniques influence predictive abilities of the models. We found that certain feature subsets show themselves as top candidates for all the models, but each model performs differently with them. We estimated models performance by ROC analysis and metrics, such as max accuracy, area under the ROC curve, area under the convex hull, partial area under the ROC curve with sensitivity above 90%, and specificity at 98% sensitivity. We paid particular attention to the metrics with higher specificity as it reduces false positive predictions, which would allow decreasing unnecessary benign breast biopsies while minimizing the number of delayed breast cancer diagnoses. In order to validate our experiments we used 5-fold cross validation. In conclusion, out results show that among the neural networks considered here, best overall performer is the Default ARTMAP neural network.

Keywords: data mining, neural networks, heterogeneous data; breast cancer diagnosis, computer aided diagnosis.

ACM Classification Keywords: I.5.1- Computing Methodologies - Pattern Recognition – Models - Neural Nets

Link:

PERFORMANCE OF COMPUTER-AIDED DIAGNOSIS TECHNIQUES IN INTERPRETATION OF BREAST LESION DATA

Anatoli Nachev, Mairead Hogan, Borislav Stoyanov

http://foibg.com/ibs_isc/ibs-23/ibs-23-p23.pdf