This study proposes a novel semi-supervised learning framework for mutually accelerated multi-contrast MRI synthesis that recovers high-quality images without demanding large training sets of costly fully-sampled source or ground-truth target images. The proposed method presents a selective loss function expressed only on a subset of the acquired k-space coefficients and further leverages randomized sampling patterns across training subjects to effectively learn relationships among acquired and nonacquired k-space coefficients at all locations. Comprehensive experiments performed on multi-contrast brain images clearly demonstrate that the proposed method maintains equivalent performance to the gold-standard method based on fully-supervised training while alleviating undesirable reliance of the current synthesis methods on large-scale fully-sampled MRI acquisitions.