3D patient motion estimation and correction for dynamic reconstruction in medical tomosynthesis

In recent years, Digital Breast Tomosynthesis (DBT) has become an increasingly important imaging modality in breast cancer detection strategies. It provides 3D information, making it more sensitive and accurate than conventional 2D mammography, while limiting patient exposure to radiations. Nevertheless, when optimal acquisition conditions, such as the absence of motion during the examination, are not met, artifacts compromising the conspicuity of clinical signs appear in the volumes, reducing the effectiveness of screening and diagnostic. This work focuses on the correction of artifacts associated with involuntary patient motion during DBT, with the aim of providing an algorithmic solution that requires no modification of either clinical protocols or imaging equipment. To this end, we propose a method based on Projection-based Digital Volume Correlation (P-DVC) to estimate --- between each projection instant --- the displacement fields corresponding to patient motion. These can then be taken into account in a further reconstruction step to reduce or eliminate artifacts. The challenge, compared to the regular use case of P-DVC, is to account for the ill-posed nature of the tomosynthesis reconstruction problem, which exploits very few projections, limited gantry angle and produces highly anisotropic volumes. In this estimation, we assume that the inconsistency between the acquired projections and the digital reprojections of the reconstructed volume (projection residual) can be interpreted mainly as resulting from the patient's displacement. To ensure this, strategies (common or original) are introduced to minimize the impact of other artifacts (truncation, background uniformity) on the minimization problem. At the same time, the richness of the kinematic basis describing all the motions that can be corrected is of major importance. It is proposed to use a progressive enrichment, using an adapted regularization strategy, while monitoring the evolution of the metrics characterizing the quality of the solution, particularly focusing on the projection residual norm and the contrast around clinical features. The validation process is based on three experimental cases, ensuring first the robustness of the solution, then its effectiveness in the clinical context. To this end, i) numerical experiments are carried out on Shepp and Logan's cranial reconstruction phantom, to which a simple perturbation is imposed during data generation. Then, ii) the process undergoes the image chain test in experiments involving a physical breast imaging phantom. Generating data from this object makes it possible to increase the complexity of the considered motions during correction, while enriching the texture and ensuring reliable knowledge of the ground truth. Finally, iii) we assess the clinical impact by applying local and global corrections to a database of breast tomosyntheses. By focusing on specific clinical signs (such as microcalcifications), a comparative assessment of image quality was carried out by clinical experts. In summary, we have shown in this work that P-DVC is applicable to an imaging modality providing very few projections and with low spatial resolution, by means of a drastic reduction in the number of degrees of freedom making up the kinematic basis. It provides high-quality volumes characterized by lower residuals, better contrast on clinically relevant signs, and displacement field estimation with voxel-level error compared to manual measurements on clinical landmarks. In addition, the implementation of a local strategy provides quick results focused on areas of interest while limiting the perturbation from surrounding tissue and computational requirements. Looking ahead, it will be interesting to further develop the use of finite element meshes to refine the dynamic reconstruction strategy for the whole breast. A statistical study could then determine an optimal reduced basis for describing patient kinematics.