A General Detector-Free Feature Matching Method for Medical Images

With the growing demand for high-precision feature matching in image-guided precision medicine, traditional methods struggle with accuracy and robustness when handling diverse medical images characterized by limited texture, significant deformation, low signal-to-noise ratio, or multi-modality. Therefore, we propose a novel general detector-free feature point matching method designed for high-precision applications in diverse medical imaging scenarios. The method builds upon LoFTR and GeoFormer within a coarse-to-fine matching framework and incorporates a cross-space multi-scale attention mechanism to enhance robustness in low-texture and noisy images across varied clinical tasks. The method consists of four main components: feature extraction, coarse-level matching, geometric enhancement module, and fine-level matching. Furthermore, we design a multi-scale matching loss function within a self-supervised learning mechanism to improve training efficiency and matching performance with limited annotated medical image data. Experimental results demonstrate that the proposed method outperforms state-of-the-art methods on retinal fundus, cervical X-ray, and pelvic X-ray datasets, exhibiting enhanced robustness in complex scenarios and low-texture regions. Additionally, we extend the method to the 2D-3D feature matching task by matching real X-ray images with multi-view DRR images generated using the DeepDRR technique, demonstrating its effectiveness and potential in cross-modal scenarios.