Automatic and reliable segmentation for geographic atrophy in spectral-domain optical coherence tomography (SD-OCT) images is a challenging task. To develop an effective segmentation method, a two-stage deep learning framework based on an auto-encoder is proposed. Firstly, the axial data of cross-section images were used as samples instead of the projection images of SD-OCT images. Next, a two-stage learning model that includes offline-learning and self-learning was designed based on a stacked sparse auto-encoder to obtain deep discriminative representations. Finally, a fusion strategy was used to refine the segmentation results based on the two-stage learning results. The proposed method was evaluated on two datasets consisting of 55 and 56 cubes, respectively. For the first dataset, our method obtained a mean overlap ratio (OR) of 89.85 ± 6.35% and an absolute area difference (AAD) of 4.79 ± 7.16%. For the second dataset, the mean OR and AAD were 84.48 ± 11.98%, 11.09 ± 13.61%, respectively. Compared with the state-of-the-art algorithms, experiments indicate that the proposed algorithm can provide more accurate segmentation results on these two datasets without using retinal layer segmentation.