When a structural brain network is constructed by the existing parcellation methods such as the automated anatomical labeling (AAL), the topological properties of the network change depending on the scale of parcellation and arbitrary connectivity matrix thresholding. To overcome these issues, we propose the ε-neighbor method, which is a parcellation free network construction technique for diffusion tensor images (DTI). The method iteratively builds network nodes and edges by adding one white matter fiber tract at a time to the current network starting from the longest tract. We examined the differences in various topological properties of the brain networks constructed using the ε-neighbor method and the existing parcellation. As the number of network nodes increased, the connectedness of the network decreased in the parcellation method. However, for brain networks constructed using the ε-neighbor method, connectedness remained at a high level even with an increase in the number of nodes. Thus, the brain networks constructed using the proposed method are considered as biologically more realistic with respect to the stability of connectedness. The proposed method is used in differentiating diffusion tensor images of normal controls and subjects with autism using various group theory features. We found significant group differences in several topological properties of brain networks constructed using the proposed method, whereas topological properties were not significantly different for the parcellation method. These results demonstrate that the ε-neighbor method overcomes the shortcomings of the existing parcellation methods and provides biologically more realistic and sensitive structural brain networks.