• class activation maps
• class distribution
• muscle segmentation
• preprocessing
• exlainability
• optimization
• breast density classifier
• mammography
• deep learning
• residual convolutional neural network

This master thesis deals with the optimization and explainability of a deep Residual Convolutional Neural Network classifier developed to assess breast density, defined as the amount of fibroglandular tissue compared to fat tissue visible on a digital mammogram. It is a risk factor for breast cancer and a parameter on which the dosimetric index depends on. An algorithm, based on deep learning methods, has been already developed to automatically classify mammograms into the 4 density classes reported in BIRADS atlas. A major problem of Deep Neural Network is their lack of transparency due to their deep multi-layer nonlinear structure. In the medical field, assessing trust in the model is fundamental for a potential application in clinical practice. The concept of explainability consists in having a look into a black box-like network to make clear the reasons behind predictions and understand its behavior. There is no well-established method and a study on explainability is completely missing in similar works, therefore two possible ways have been explored. Primarily, the classifier has been trained in different conditions to study how the output varies with the input. In this phase, preprocessing step, model architecture and class distribution in the dataset have been taken into account as factors that influence the classifier performance. Then, off-line visualization techniques have been used. Class Activation Maps, i.e. images that highlight the regions of the mammogram on which the attention of the algorithm is focused on, have been generated with a gradient-based method and qualitatively evaluated. These analyses have led to a better understanding of how the algorithm works and to an improvement in its classification performance in terms of accuracy.