Better utilizing the vast amount of valuable information stored in the medical imaging databases is always an interesting research area, and one way is to retrieve similar images as a reference dataset to assist the diagnosis. Distance metric is a core component in image retrieval; and in this paper, we propose a new learning-based distance metric design, based on regression and classification techniques. We design a weight learning approach by classifying the similar-dissimilar data samples, and a further optimization with a sparsity-constraint regression algorithm for feature selection. The learned distance metric is generally applicable for medical image retrievals. We evaluate the proposed method on clinical PET-CT images, and demonstrate clear performance improvements.