The invention discloses a radiomics feature selection method based on momentum adaptive harmony search, and belongs to the technical field of radiomics. The method comprises the following steps of: 1, checking a region of interest (ROI) of the medical image; 2, adrenal tumor CT image data containing the ROI are read; 3, extracting original image features from the read image data; 4, defining an objective function of the method for minimizing the problem; 5, initializing a harmony memory bank; 6, initializing the maximum number of iterations, the HMCR and the PAR; 7, generating a new solution j; 8, updating the harmony memory bank; and 9, repeating the steps 7-8, and outputting an optimal feature subset. According to the method, the HMCR and the PAR are dynamically adjusted, so that sinking into a local optimal solution is avoided, and the robustness of the algorithm is enhanced; the concept of momentum gradient descent is fused, and convergence is accelerated; the fitness function is adjusted, and the essence of the feature search optimization problem is better conformed.