Brain disease is a top cause of death. Currently, its main diagonosis is to take advantage of medical brain images to analyse patients’ condition. In medical big data analysis field, it has been a research hotspot that how to effectively represent medical images and discover significant information hidden in them to further assist doctors to achieve a better diagnosis. Graphs, as one of the most general forms of data representation, can easily represent entities, their attributes and their relationships well. However, the existing medical image graph models do not exploit the specific relationships of brain images very well so that some essential information is lost. Therefore, aiming at brain images, we firstly construct a domain knowledge-oriented graph about the Topological Relationships among Ventricles and Lesions (TRVL) to represent a brain image, and give the algorithm of modeling a brain Image to a TRVL Graph (denoted as I2G). Then we propose a method named Frequent Approximate Subgraph Mining based on Graph Edit Distance (FASMGED) to exactly discover meaningful patterns hidden in brain images. This method employs a strong error-tolerant graph matching strategy which is accordant with ubiquitous noise in practice. Moreover, an approximate method of frequent approximate subgraph mining is proposed based on the greedy strategy. We have evaluated our algorithms on real and simulated data. Results show that I2G is computationally scalable, FASMGED can discover more significant patterns than other state-of-the-art frequent subgraph mining methods, and the approximate method of frequent approximate subgraph mining outperforms FASMGED.