Neuroimaging is an important research platform that can be very useful for eliciting new understanding on the complicated pathogenesis between genetics and disease phenotypes. Due to the extremely high dimensionality of image and genetic data, and considering the potential joint effect of genetic variants, multivariate techniques have been examined to detect Alzheimers disease (AD) related genetic variants expressed through single-nucleotide polymorphisms (SNPs). However, the image features used in support of those methods are not immediately related to the disease, and the detected genetic markers may not be related to AD. In this study, we propose an ensemble model based framework for firstly extracting 50 region-based image features whose values are predicted by base learners trained on raw neuroimaging morphological variables. This task is followed by performing sparse Partial Least Squares regression (sPLS) method on the extracted 50 AD related image features and pre-selected 1508 SNPs to detect the significant SNPs associated with the extracted image features. Instead of modeling a direct link between genetic variants and disease label, we captured disease information indirectly.