Comparison of image registration methods for combining laparoscopic video and spectral image data

Laparoscopic procedures can be assisted by intraoperative modalities, such as quantitative perfusion imaging based on fluorescence or hyperspectral data. If these modalities are not available at video frame rate, fast image registration is needed for the visualization in augmented reality. Three feature-based algorithms and one pre-trained deep homography neural network (DH-NN) were tested for single and multi-homography estimation. Fine-tuning was used to bridge the domain gap of the DH-NN for non-rigid registration of laparoscopic images. The methods were validated on two datasets: an open-source record of 750 manually annotated laparoscopic images, presented in this work, and in-vivo data from a novel laparoscopic hyperspectral imaging system. All feature-based single homography methods outperformed the fine-tuned DH-NN in terms of reprojection error, Structural Similarity Index Measure, and processing time. The feature detector and descriptor ORB1000 enabled video-rate registration of laparoscopic images on standard hardware with submillimeter accuracy.