Convolutional Neural Network (CNN) based classifier for breast density assessment

Breast cancer is one of the most diagnosed cancer all over the world. It has been studied that one woman in eight is going to develop a breast cancer in her life. It is also widely accepted that early diagnosis is one of the most powerful instrument we have in fighting this sort of cancer. For these reasons, in Tuscany, mammographic screening programs are performed on asymptomatic women at risk every two years in a range between 45 and 74 years. Full Field Digital Mammography (FFDM) is a non-invasive high sensitive method for early stage breast cancer detection and diagnosis, and represents the reference imaging technique to explore the breast in a complete way. Since mammography is a 2D x-ray imaging technique, it suffers of some intrinsic problems: a) breast structures overlapping, b) malignant masses absorb x-rays similarly to the benignant ones and c) the sensitivity is lower for masses or microcalcifications clusters in denser breasts. In fact, a mammogram with a very high percentage of fibro-glandular tissue is less readable because dense tissue presents an x-ray absorption coefficient similar to cancer’s one. Furthermore, to have a sufficient sensitivity in dense breast, a higher dose is given to the patient. Since a lot of healthy women are called to partecipate to the screening programs, dose delivering should be carefully controlled. Furthermore, the European Directive 59/2013/EURATOM states that patients must be well informed about the amount of received radiation dose. For these reasons, the RADIOMA project (RADiazioni IOnizzanti in MAmmografia) was born with the aim of developing a personalized and reliable dosimetric quantitative index for mammographic examinations. The purpose of this master thesis was to build a breast density classifier in order to personalize the new dosimetric index according to breast density. Since the most used density standard has been established by the American College of Radiology (ACR) in 2013, we decided to use those classes to train the classifier. This standard is written on the Breast Imaging Reporting and Data System (BIRADS) Atlas and it is made of four qualitative classes: almost entirely fatty (“A”), scattered areas of fibroglandular density (“B”), heterogeneously dense (“C”) and extremely dense (“D”). In this master thesis, a deep learning based technique has been explored to build the classifier. In fact, in the last few years, deep learning-based methods have been developed with success in a wide range of medical image analysis problems. Since deep learning needs a huge amount of data, the “Azienda Ospedaliero-Universitaria Pisana” (AOUP) collected about 2000 exams from the Senology Department. The exams has been selected by a mammography specialized physician and a radiology technician. This dataset has been anonymized and extracted from the AOUP database. Once obtained the dataset, first, I tried to solve a preliminary classification problem using only two of the four BIRADS classes: the A class, made of less dense breasts, and D class, made of densest breasts. The chosen architecture to solve this problem is a VGG architecture, in which several convolutional layers are stacked together in order to have a high number of input image representations, keeping low the number of parameters. After obtaining good results from this classifier, I proceed to build a more complex classifier. Since one of the main problem in mammograms classification is to build a classifier able to discriminate between dense and non-dense breast, I trained a CNN to solve this problem. In fact, some clinical decisions depend on the possible masking effect that dense tissue could produce on a mammogram. In BIRADS density standard, this means that we should classify two classes: the first one is made of mammograms belonging to A and B classes and the second one is made of mammograms belonging to C and D classes. To solve this problem, I chose a residual architecture, in which the network is asked to learn the residual mapping of convolutional layers. The variability of test accuracy, over input size and over the number of projections used, has been studied. The highest test accuracy is equal to 89.4% and this means that the classifier is able to predict the rightlabel of new mammograms with a precision equal to 89.4%. Finally, I trained a Convolutional Neural Network with a very deep residual architecture to build a BIRADS classifier. As first step, the best optimization of the hyperparameters have been performed. Afterwards, a sistematic study of accuracy variation over input size and the projection used has been performed. The highest test accuracy is equal to 78.0%. Both dense/non-dense and BIRADS classifier have been built with 4 CNNs, one for each projection. The last layer of each CNNs, which represents the score classification, has been averaged over the four mammographic projection in order to assess density as a overall evaluation on an entire mammographic exam, as radiologists do. For the BIRADS classifier, a further rule to produce the final label of the test set has been established: since a radiologist assigns the highest density class when a breast density asymmetry occurs between right and left breasts, in this work, the final label has been assessed separatly for right and left breasts and then the highest density class is assigned to the patient. Regarding the first problem, the convolutional neural network trained on 650x650 pixels images predicts the right label with an accuracy equal to 89.4%. This classifier works well at least as much as other classifier described in other previous works. Regarding the BIRADS classification, I obtained a test accuracy on 650x650 pixels images equal to 78.0%. This result is very good compared to test ac- curacy of other classifier of breast density. Some achievable improvements can be performed in order to have a higher accuracy and a better generalization. First, the ground truth of this work has been decided by only one radiologist. Since the intra-observer and inter-observer variabilities are quite high in BIRADS classification, we could produce a ground truth using the maximum agreement between more than one radiologist. Second, we can increase the number of exams in training, validation and test set in several ways. In fact, other clinical mammograms are going to be collected and, thanks to a recently born collaboration between RADIOMA project and “Azienda Toscana Nord-Ovest” (ATNO), screening exams are going to be collected too. Furthermore, if we find a well-working image standardization method, we may be able to analyze together mammographic exams obtained with different mammographic imaging systems. Finally, using more powerful GPUs, we may be able to train the CNN with high resolution images and with the four mammographic projection at the same time. Beyond this work, we can exploit such technique to find new relationships between mammograms and known breast cancer risk factors. Since breast density is a well known risk factor, which is not considered in the most used breast cancer risk models, CNNs can be used to assess quantitavely and automatically not only breast density but also its role in cancer developing