Identification of mammography anomalies for breast cancer detection by an ensemble of classification models based on artificial immune system

• Published in: Knowledge-based systems Vol. 101; pp. 60 - 70
• Main Authors: Magna, Gabriele, Casti, Paola, Jayaraman, Sowmya Velappa, Salmeri, Marcello, Mencattini, Arianna, Martinelli, Eugenio, Natale, Corrado Di
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-06-2016
• Subjects: Accuracy; Algorithms; Artificial immune system; Artificial intelligence; Breast; Breast cancer; Cancer; Classification; Ensemble of classifiers; Mathematical models; Ensemble of classifiers; Breast cancer; Artificial immune system
• DOI: 10.1016/j.knosys.2016.02.019
• EISSN: 1872-7409
• ISSN: 0950-7051

The interpretation of diagnostic images is often conditioned by the specific properties of the instrument that generated the image. This makes particularly complicated to develop universal recognition algorithms that can facilitate the diagnosis in case of massive population screenings. Mammography is a typical example where such an algorithm is required. Although the technological advances in medical imaging increases the accuracy of interpretation of images, the improved resolution may not facilitate the identification of breast cancer at a very early stage due to the many confounding factors related, for instance, to differences in instrument settings or breast positioning by the operator. Being impossible an exact standardization, the problem of reducing the effects of different instruments and operators can be faced with a proper algorithm that can extract from each image the relevant information in an unsupervised manner, thus limiting the influence of instrumental and positioning issues. For this scope, in this paper we investigated the properties of a classifier based on an ensemble of Adaptive Artificial Immune Networks (A2INET) applied to original mammography image indicators aimed at diagnosing bilateral asymmetry that is known to be correlated with increased breast cancer risk. Classification models were trained using a set of descriptors measuring the degree of similarity of paired regions of the left and right breasts. Noteworthy, the ensemble of A2INET models achieved very high classification rates even when training and testing were made on two completely independent and heterogeneous datasets. The obtained results are promising (maximum accuracy level of 0.90, sensitivity of 0.93 and specificity of 0.87) and they prefigure to apply automatic diagnostic tools in clinical practice exploiting a network of different instrument databases.