While convolutional neural networks have shown promise in medical image registration, their inherent complexity limits their registration speed, particularly for surgical applications. Additionally, traditional feature-based matching methods struggle with multi-modal forearm image registration due to the simplicity of forearm skin textures. To address these issues, we propose a robust forearm feature point extraction method based on the forearm’s structural invariance. We combine this method with thin plate spline interpolation to achieve multi-modal forearm registration. Our approach introduces the Forearm Feature Representation Curve (FFRC) and the Multi-Modal Image Registration Framework (FAM) for aligning forearm images with digital anatomical models. FFRC identifies feature points based on forearm structural characteristics, and FAM employs FFRC for matching point pre-screening before applying an affine transformation. For deformable registration which adds Thin Plate Spline (FAM-TPS) uses the matched points as control points. In our experiments, both FAM and FAM-TPS demonstrate high registration accuracy, with FAM-TPS outperforming conventional feature-based methods. Our framework excels at registering forearm images with varying rotation angles, and we have observed a strong correlation between the feature curve’s peak value and the rotation angle. These results affirm the effectiveness of our approach in achieving precise and resilient registration.
|