In this paper, we address the challenge of rotation-invariant feature extraction in image processing, and propose a rotation-invariant feature extraction method based on Radon projections. According to the principle of the Fourier Central Slice Theorem, our method employs the inverse process of Computed Tomography (CT) reconstruction to efficiently obtain multi-angle Radon projections of the input image. After this transformation, rotating an image results in cyclic shifts of the projections, which essentially changes the order of the angle sequence. Subsequently, these projection sequences from different angles are processed through Multi-Head Self-Attention (MHSA) module, which integrates the correlations of the image from different viewing perspectives into rotation-invariant features. Experimental results demonstrate that our method enhances the performance in rotation tests.