The application of integrative regression analysis using neuroimaging data from diverse sources has been crucial in investigating the relationship between clinical factors and brain structures, thereby aiding in the comprehension of the mechanisms underlying neurological diseases such as Alzheimer’s disease (AD). However, it usually suffers from imaging heterogeneities caused by differences in study design, protocol, environment, population, or other hidden confounders. To address this challenge, this paper introduces a novel model known as the deep image-on-scalar regression model with hidden confounders (DISRM-HC), which can simultaneously detect hidden confounders and learn primary effects from variables of interest by employing the functional surrogate variable analysis and deep neural networks. Compared to existing solutions, our DISRM-HC can successfully handle imaging heterogeneities and capture complex patterns of primary effects. In our model, we establish estimation procedures for both unknown varying coefficients and hidden confounders. The asymptotic properties of the estimation procedures are systematically investigated as well. Subsequently, the finite-sample performance of our proposed DISRM-HC is assessed using synthetic datasets generated from Monte Carlo simulations and a real data example on diffusion MRI images from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) study.