An Optimized Feed-forward Artificial Neural Network Topology to Support Radiologists in Breast Lesions Classification

Introduction and objective: Computer Aided Decision (CAD) systems based on Medical Imaging could support radiologists in classifying malignant regions from benign ones, in the field of investigation for breast cancer detection. This decision may often follow a previous procedure dedicated to the earlier identification of Regions Of Interest (ROI) containing still unclassified lesions. Materials and methods: Materials comprise features extracted from magnetic resonance (MR) images representing morphological properties of lesions. The Regions Of Interest identified by a previous automatic procedure validated by radiologists of the University of Bari Aldo Moro (Italy), authors of this work, 134 from 600 slices considered of interest, because they contain still unclassified damaged areas. Several techniques were tested for ROI segmentation and classification. In particular, it can be shown that the same procedures for lesioned-area discrimination were also useful for malignancy classification of lesions, themselves. In particular, MR images were processed with different image processing techniques for ROI extraction, which were, ultimately, described by morphological features, such as circularity, aspect ratio, solidity and convexity. Finally, we discuss a procedure to design a feed-forward supervised artificial neural networks (ANN) architecture based on an evolutionary strategy. In a similar approach, different ANN topologies were tested in order to find the best in terms of mean accuracy for several iterations of training, validation and test. In particular, for each topology, the training, validation and test sets were constructed using 100 random permutations of the dataset, from which the average performances were calculated. Results: The performance of the best ANN architecture, trained using a training set of 82 samples (equally divided between malignant and benign lesions) from the 134 samples available in the whole dataset, were evaluated in terms of accuracy, sensitivity and specificity. Conclusion: Testing determined that the supervised ANN approach is consistent and reveals good performance; in particular, the optimal ANN topology found through an evolutionary strategy showed high generalization on the mean performance indexes regardless of training, validation and test sets applied, showing good performances in terms of both accuracy and sensitivity, permitting correct classification of the true malignant lesions.