We consider in this thesis the statistical linear inverse problem $Y = Af+ \epsilon \xi$ where $A$ denotes a compact operator, $\epsilon>0$ a noise level and $\xi$ a Gaussian white noise. The unknown function f has to be recovered from the indirect measurement Y . Given a family $\Lambda$, an oracle inequality compares the performances of an adaptive estimator $f^{\star}$ to the best one in $\Lambda$. Such an inequality is non-asymptotic and no specific informations on $f$ are required. In this thesis, we propose different oracle inequalities in order to provide both a better understanding of regularization with a noisy operator and ageneralization of the risk hull minimization (RHM) algorithm. For most of the existing methods, the operator...