Journal of Digital Information Management J. Wang, S.Y. Lai, M.D. Li, D. H. Lies 13 6 2015 http://dline.info/fpaper/jdim/v13i6/v13i6_3.pdf A gradient-projected relevance feedback algorithm based on non-negative matrix factorization (NMF) is proposed in this study to improve the performance of retrieval algorithm in the medical image processing field. Relevance feedback has been an important method in image retrieval technology in recent years because it allows users to participate. Thus, it can compensate for the shortcomings of using low-level features to describe the semantic contents of an image to some degree. Given that NMF can partly sketch the distribution of relevant images in the space represented by the base matrix, finding more related images from image repositories is possible. This condition can be achieved by conducting an NMF operation of the query image, using the gradient projection iterative rules to update variables, and selecting the appropriate iteration stop conditions to optimize the time complexity of the algorithm. Compared with the commonly used and multiplicative updating NMF approaches, the proposed method improved the speed of the feedback on the premise of guaranteeing precision and recall rates, and significantly optimized the retrieval accuracy. Experiments were conducted on the base of 586 cerebral hemorrhage images and 634 spine and cervical-spine mixed images. Results show that the proposed approach is feasible in medical image retrieval.