Machine learning has been widely adopted to medical imaging research, yet it suffers from domain shift for real world applications. Due to the heterogeneity of medical data, machine learning-based diagnostic models are also prone to biases. In this thesis, we start from arguing the necessity of domain adaptation to achieve the optimal performance for each subcategory. We develop an adaptation algorithm which doesn’t require any ground-truth labels from the unseen domain. We also discuss the value of handcrafted imaging features in the representation learning era for brain imaging application. Next, we show that machine learning-based diagnostic models can be unbiased if they are trained using rigorous data pre-processing techniques and well-constructed models. We find that multi-source data is helpful sometimes in elevating both model performance consistency and precision. Finally, we introduce a weighted-empirical risk minimization algorithm to further boost the model’s performance on unseen data distribution by only using a few samples. We demonstrate the empirical results on large-scale diverse-population brain imaging datasets for rich clinical tasks.