Robust X-Ray Image Stitching Algorithm Based on Refining Matching Results of Feature Descriptors

Panoramic X-ray image provides a convenient way for some orthopedic clinical diagnosis and preoperative planning, while the full body cannot be captured in a single X-ray scan. The classic refining algorithm RANSAC failed to identify the outliers (mismatched points) according to the low signal-to-noise ratio of the matching results of vanilla feature descriptors. The methods based on deep learning based are also difficult to carry out due to the lack of training data pairs clinically. This paper proposes a robust algorithm for X-ray image stitching based on refining the matching results of vanilla feature descriptors. The algorithm transforms the image stitching task into the clique problem in the graph theory owing to the scale consistency of the X-ray images, and the dynamic programming algorithm is used to accelerate the searching process for the maximal clique. 30 clinical X-ray image pairs from different parts of the human body are randomly collected to evaluate our proposed algorithm. The experiment results show the strong robustness throughout different stitching tasks of X-ray images and achieve the best performance among the frequently-used stitching software.