Multi-view convolutional neural networks for breast cancer diagnosis

The breast cancer classification task is a widely tackled problem in deep learning, since the computational power of computers has been allowing the analysis of medical images for several years now. Although new types of neural network architectures, such as Visual Transformers and Anatomy aware Graph convolutional Networks, have recently emerged and seem to be promising, the majority of the solutions proposed in the literature are still based on Convolutional Neural Networks, which are known to be efficient for computer vision applications. In order to compare the performances of the three types of architectures mentioned, three models, each based on one of the architectures, have been trained and evaluated on the same dataset, and the same classification tasks. This thesis is dedicated to the convolutional-based standard approach, and is part of a global research project aiming to compare the three different models in order to implement a triage system. The standard model studied in this thesis is the image-only view-wise model that was originally developed by researchers from New York University in 2020. This model bases its predictions on the simultaneous analysis of the 4 views of a mammogram, reproducing therefore the behavior of a human radiologist. Initially trained and evaluated on over 200,000 exams (over 1,000,000 images), the New York University researchers reported an AUC score of 0.687 +/- 0.009 for the malignant finding classification task. The version of the view-wise model I implemented was slightly modified and adapted to our common dataset and classification tasks, which are the detection of a malignant tumor (primary task) and the recall of an exam (secondary task), both for the left and right breasts. The dataset I used for the training and evaluation was obtained from the cohort of the screen-aged woman case-control subset, acquired in the region of Stockholm, and counts, in comparison, only 17,380 exams. It also reveals a large disproportion between controls and cases, as it reflects a screening population. Training a model for such a complex task, with so little data is a real challenge. In order to counterbalance these problems, I show that the increasing of the training set by adding 1,952 exams from the CBIS-DDSM public dataset, enhanced the performances of my model on the validation set, despite the differences between the two datasets. Furthermore, I proposed a methodology to generate additional synthetic positive examples, by extracting lesions from annotated cases and blending them on healthy controls, adding therefore another 1,112 extra cases (5 times the initial number of cases). This technique has contributed to increase the number of samples, without requiring external data. Finally, by designing an adequate sampling strategy for the training set, ensuring that the network is seeing enough positive examples at each epoch, in order to learn correctly its classification task, I was able to observe comparable results with respect to those reported in the New York University paper. All those techniques appear to be particularly relevant and helpful to improve the training of a model, especially in a context where the data are limited and difficult to obtain.