CIFRE PhD Thesis under the supervision of Prof. Nicolas Thome, in collaboration with GE Healthcare. Subject: Automatic image analysis and synthesis for optimizing clinal performances in 3D mammography. Mammography is known, through several mass screening trials, for its ability to reduce the breast cancer mortality rate for women. Over the last 15 years, mammography has entered the digital imaging era, making it possible to perform screening, diagnostic, and interventional procedures more effectively. In particular, we have seen the development of Digital Breast Tomosynthesis (DBT), a significant advance in mammography technology to enable "semi-3D" mammogram visualization. With the success of Deep Learning (DL) for visual recognition, especially the great performance obtained by Convolutional Neural Networks (ConvNets) on the ILSVRC challenge in 2012, its application to mammography data has seen tremendous growth in popularity. In this thesis, we are interested in exploring and developing methods for automatic analysis and lesion detection in DBT. Our study focuses on two main questions: learning based on limited examples, and localizing lesions in 2D mammograms and 3D DBT.
- Multitask Classification and Segmentation for Cancer Diagnosis in Mammography. , Annotation cost is a bottleneck for collecting massive data in mammography, especially for training deep neural networks. In this paper, we study the use of heterogeneous levels of annotation granularity to improve predictive performances. More precisely, we introduce a multi-task learning scheme for training convolutional neural network (ConvNets), which combines segmentation and classification, using image-level and pixel-level annotations. In this way, different objectives can be used to regularize training by sharing intermediate deep representations. Successful experiments are carried out on the Digital Database of Screening Mammography (DDSM) to validate the relevance of the proposed approach. www