Rôle de l'apprentissage profond dans la reconstruction tridimensionnelle en tomosynthèse du sein

Digital breast tomosynthesis (DBT) is an X-ray imaging technique introduced in the 2010s that provides a three-dimensional reconstruction of tissues. Commercial systems rely on acquiring images from a limited number of views sampled every 1 to 3 degrees, evenly distributed over a scanning cone of 15 to 40°. The acquired projections are then processed by a reconstruction algorithm that generates the volume reviewed by the radiologist. Numerous analytical and neural network reconstruction algorithms exist. However, the narrow opening of the acquisition cone limits the quality of the reconstructed volume, particularly its resolution in the thickness of the breast. Orthogonal slices thus become unreadable with current techniques :the gradual extinction of artefacts from an object outside its focal planes disrupts the reading of objects located in neighbouring planes. Thus, there is currently no commercial solution effectively addressing limited-angle artefacts in breast tomosynthesis.The use of deep learning faces several challenges. Indeed, paired data in real conditions that would allow supervised learning are absent,and volumes from other breast imaging modalities cannot be used directly . Questions also arise regarding the properties of the obtained methods. On the one hand, the reconstructed volumes are not always consistent to the measurements, a property well mastered by conventional methods. Moreover, the uncertainty of the image reconstructed by deep learning is almost never explicit. In this thesis, we first propose a post-processing approach for conventional reconstruction. To develop our method, we create a synthetic database consisting of digital phantoms. We simulate their X-ray projections and reconstruct 3D volumes using conventional methods,thus creating a dataset that we use to train a convolutional neural network in a supervised manner. We demonstrate that the proposed strategy significantly improves the quality of orthogonal planes and is therefore promising to address the problem of reconstruction in breast tomosynthesis. Secondly, we seek to obtain an estimate of the reliability of the volumes predicted by the neural network. We adopt a Bayesian perspective, differentiating between aleatory and epistemic uncertainties. We model the first term using a Laplace distribution and the second by approximating the posterior predictive distribution, then compare the results obtained with a Monte Carlo Dropout method and a deep ensemble. We show that the computed uncertainty is a good approximation of the actual error and reuse to minimise a data consistency term. Finally, we improve the realism of the synthetic volumes by adapting images from another imaging modality that does not present the same artefacts as tomosynthesis. After appropriate denoising, we segment these images and simulate their compression to replicate the conditions of the medical examination. We redesign the neural network to allow 3D learning that considers information from the three anatomical directions, and solve the convergence problems related to the calculation of the uncertainty terms. We then impose the fidelity constraint to the projections by limiting the contribution of the neural model to the kernel space of the projection operator. Finally, we evaluate our methods on volumes with specific geometric properties as well as on clinical images, and highlight the benefits and limitations of our deep learning reconstruction approaches for breast tomosynthesis.