Neuroimaging made it possible to quantify brain structure and function. However, there are few neuroimaging biomarkers for the early diagnosis, prognosis, and evaluation of therapy for brain diseases. The development of neuroimaging biomarkers for brain diseases faces two major bottleneck problems. First, the neuroimaging datasets of brain diseases are always characterized by small sample size, high dimension, and large heterogeneity. Second, a fine-grained individualized human brain atlas for effective dimensionality reduction has always been lacking. Due to the inherent high-dimensional nature of neuroimaging, collecting a large amount of data on specific brain diseases of interest is essential to avoid model overfitting, obtain meaningful biological insights, and enhance statistical power. However, limited by real-world clinical scenarios, it is often difficult for a single center to obtain massive samples of specific diseases. So multi-center data analysis has gradually become a research trend. Besides, on the one hand, a reliable brain atlas can be used as prior knowledge to effectively reduce the dimensionality of neuroimaging to reduce the possibility of overfitting. On the other hand, it also helps us to integrate research conclusions, to further reveal the mechanism of the disease. Therefore, a systematic framework to accurately locate brain regions and networks, and then investigate how to effectively aggregate multi-center neuroimaging datasets and robustly search imaging features for the comparison of different individuals is needed. This will help to clarify the imaging biomarkers of brain diseases and push neuroimaging into clinical applications.