Signal-dropouts caused by cardiac motion are frequently observed artifacts in diffusion-weighted imaging (DWI) of the liver. When several repetitions of a given slice are affected, uniform averaging results in locally reduced liver signal. This work proposes an adaptive weighted averaging of repetitions to locally suppress signal-dropouts. Therefore, weight maps are estimated by an algorithm which computes robust patch statistics under the guidance of a learned classifier which marks corrupted repetitions. In comparison to other methods, the proposed approach enables more homogeneous liver signal and less biased quantitative maps while sacrificing little signal-to-noise ratio (SNR).