Abstract: To improve the 3D spatial feature extraction ability of current brain disease classification methods, a novel classification method based on 3D attention convolutions and self-supervised learning was proposed. First, a 3D attention convolutional neural network based on residual connections was proposed to extract spatial features, aiming to distinguish the importance of different spatial positions in voxel data using 3D attention mechanism. Second, a multi-task learning framework based on self-supervised learning was constructed by using the learned spatial features, and a self-supervised auxiliary task based on spatial continuity was used to further mine the spatial dependence of voxels. Finally, the model parameters were optimized by jointly training the target classification task and self-supervised auxiliary task, and then the classification performance can be largely boosted. Experimental results on ABIDE-Ⅰ and ABIDE-Ⅱ show that the proposed method has superior classification performance and the results also demonstrate good interpretability.