Computer-Aided Diagnosis of Breast Masses using Shape and Margin Radiomic Descriptors and Neural Networks in Dedicated Breast CT Imaging

Objective: The goal of this Master Thesis is to design a Computer-Aided Diagnosis (CADx) system for breast mass-like lesion classification in Dedicated Breast CT (DBCT) images, using a quantitative radiomics approach based on newly developed shape and margin imaging biomarkers and a multi-layer perceptron neural network (NN). The clinical motivation behind it is to reduce the number of negative and unnecessary breast biopsies, which constitutes more than 70% of the overall biopsies performed. This project was carried out at the Advanced X-ray Tomographic Imaging (AXTI) Laboratory, Department of Radiology and Nuclear Medicine, Radboud University Medical Center (Nijmegen, The Netherlands). Methods: A traditional radiomic pipeline was implemented. Therefore, starting from DBCT images and their manual segmentation, the main phases performed were the image cropping to obtain the patches containing the breast lesions, the data augmentation process to increase the number of available patches, the feature extraction based on shape and margin descriptors, and the implementation of the diagnostic Machine Learning (ML) model for the classification of benign and malignant breast masses. The images used in this study were acquired with the new DBCT imaging modality and consisted of 74 breast lesions (54 benign and 20 malignant). As concerns the breast masses analysis, this work was focused on the development, validation, and implementation of new descriptors that quantify the tumor properties in terms of morphology (shape and contour) and margin appearance (border heterogeneity and infiltration degree). The final feature set contained 158 features, some of which were already included in previously published studies, others were newly designed and implemented in this thesis. As regards the classification task, this work involved a study related to the NN architecture in terms of the number of layers and neurons (NN tuning), the number of features to consider (feature selection), and the number of samples to be used during the network training (dataset balancing) in order to correctly predict the breast lesion diagnosis. Results: The final NN model resulted in a sensitivity of 0.79, a specificity of 0.90, an F1-Score of 0.79, an accuracy of 0.86, and an AUC of 0.91. It contained 56 input features (feature selection using Random Forest), two hidden layers of 10 neurons each, and one output layer with a Sigmoid activation function. The dataset used for its training was balanced by undersampling the class of benign lesions (majority class). In particular, dataset balancing made it possible to significantly increase the performance, compared to a NN classification model based on the initial imbalanced dataset (1:3 in this study). Instead, feature selection showed that its contribution is minimal both on the imbalanced and balanced dataset, so it did not constitute a critical step, underlying the importance of these features in the discrimination between benign and malignant masses. Conclusions: The inclusion of radiomic features that assess morphology, margin and peritumoral compartments of breast lesions in the development of a CADx system for DBCT images could provide high classification performance. Therefore, their combination with the well-known and traditional texture features could lead to a radiomic analysis aimed at quantifying all the most distinctive characteristics of breast masses, leading to significant diagnostic decision support.