Fusion of 3D Radiomic Features from Multiparametric Magnetic Resonance Images for Breast Cancer Risk Classification

• Published in: Smart Technologies, Systems and Applications pp. 259 - 272
• Main Authors: Marín-Castrillón, Diana M., Rincón, Jaider Stiven, Castro-Ospina, Andrés E., Hernández, Liliana, Díaz, Gloria M.
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing
• Series: Communications in Computer and Information Science
• Subjects: Breast cancer; Fusion strategies; MRI sequences; Radiomics
• ISBN: 3030467848; 9783030467845
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-3-030-46785-2_21

Radiomics imaging technology refers to the computation of a large number of quantitative features to describe characteristics of medical images that specialists can not appreciate. The quality of Magnetic Resonance Images (MRI) allows capturing pathology features through different sequences acquired from the target tissue. The use of multiparametric MRI has shown to be useful in the diagnosis of breast cancer but is challenging to radiomic analysis by the fusion of information from different images. In this work, 3D radiomic features extracted from nine breast MRI sequences were used to train a model able to discriminate positive and negative breast masses in the region of interest manually identified. Two fusion strategies were here evaluated; in the first one, features from all sequences were concatenated, and Fisher-Score and Gini-Index feature selectors were used to identify the most discriminative features; in the second one, features were initially selected from each sequence and then were concatenated to be used for training a classification model; Random Forest and a Support Vector Machine learning models were also evaluated. To test both fusion strategies was used a database with 146 Interest volumes (VoIs) from which, 61 were positive and 85 negative findings. Although there are differences in the accuracy obtained by the two fusion strategies, these are not significant and it was clear that a small number of features provide better performance than using the whole set of them. The best performance was reached using Random Forest with the 20% of the computed features and selected by the Gini/index algorithm, With which an AUC of 73.6% was obtained.