Recently, graph anomaly detection has attracted increasing attention in data mining and machine learning communities. Apart from existing attribute anomalies, graph anomaly detection also captures suspicious topological-abnormal nodes that differ from the major counterparts. Although massive graph-based detection approaches have been proposed, most of them focus on node-level comparison while pay insufficient attention on the surrounding topology structures. Nodes with more dissimilar neighborhood substructures have more suspicious to be abnormal. To enhance the local substructure detection ability, we propose a novel Graph Anomaly Detection framework via Multi-scale Substructure Learning (GADMSL for abbreviation). Unlike previous algorithms, we manage to capture anomalous substructures where the inner similarities are relatively low in dense-connected regions. Specifically, we adopt a region proposal module to find high-density substructures in the network as suspicious regions. Their inner-node embedding similarities indicate the anomaly degree of the detected substructures. Generally, a lower degree of embedding similarities means a higher probability that the substructure contains topology anomalies. To distill better embeddings of node attributes, we further introduce a graph contrastive learning scheme, which observes attribute anomalies in the meantime. In this way, GADMSL can detect both topology and attribute anomalies. Ultimately, extensive experiments on benchmark datasets show that GADMSL greatly improves detection performance (up to 7.30% AUC and 17.46% AUPRC gains) compared to state-of-the-art attributed networks anomaly detection algorithms.