Attention deficit hyperactivity disorder (ADHD) is a common mental-health disorder in adolescent groups. Successful automatic diagnosis of ADHD based on features extracted from magnetic resonance imaging (MRI) data, would provide reference information for treating. Previous researches have shown gray matter (GM) of some anatomical brain structures will increase in ADHD subjects. Fractal analysis has been widely used in texture image processing and fractal dimension is capable of representing intrinsic structural information of images. With large-scale MRI data becoming publicly available, deep-learning methods for ADHD diagnosis become feasible. This paper proposes a novel classification approach using 3D fractal dimension complexity map (FDCM) for ADHD automatic diagnosis. We calculate the Hausdorff fractal dimension of GM density data extracted from structural MRI data. Subsequently, we design a 3 dimensional convolutional neural network (3D-CNN) for extracting features from FDCM then judging ADHD and TDC. Our model is evaluated on the hold-out testing data of the ADHD-200 global competition and performance outperforms previous approaches based on structural MRI data.