Artikel
Image mosaicing in neuroendoscopy
Image Mosaicing und virtuelle Bildausschnittserweiterung in der Neuroendoskopie
Suche in Medline nach
Autoren
Veröffentlicht: | 30. Mai 2008 |
---|
Gliederung
Text
Objective: In neuroendoscopy the surgeon has to deal with a rather limited field of view which can cause navigational difficulties. It is therefore desirable to have a tool which combines automatically many endoscopic video frames to a larger, metrically accurate field-of-view (image mosaicing).
Methods: The algorithm is based on the method of Kourogi [1] which we extend to the case of endoscopic masks. The algorithm finds fully automated the optimal affine transform between video frames and builds the enlarged field-of-view. We apply our algorithm to endoscopic video sequences and compare it to the well-known image-mosaicing algorithm of Szeliski [2]. The goal of our algorithm is to estimate the motion field between successive frames I(t-1) and I(t) of a video sequence. This is done with an improved optical flow algorithm. It calculates at each pixel (x,y) the so-called pseudo motion based on the lumincance signal / (x,y,t) in frame t. Our algorithm has the option to 'undistort' every frame. The radial distorsion parameters are known from previous camera calibration.
Results: In a first experiment we created an endoscopic video sequence (30 frames) of grey level images where each frame is connected to the next by a known affine transform. We tested the two above- mentioned algorithms applied to the same task. With our algorithm the resulting image mosaic gives a much better overview than the single frames and is free of mosaicing artefacts. In a second experiment we obtained results on a real neuroendoscopic color sequence where the true motion is not known. Although the lighting conditions change, our method is able to find very good registrations for the image mosaic. Small border artefacts are hardly visible. All images were 'undistorted' prior to image mosaicing. A parallel run of the same experiment on the original distorted images resulted in an inferior mosaic, indicated by a decline of 4% in the rate of accepted pixels. We note that our method is by a factor of 3 faster than the algorithm of Szeliski. An implementation of our algorithm as Java-based ImageJ-plugin runs at a speed of 5-12 fps (frames per second) on a standard 1.5 MHz Pentium PC.
Conclusions: We have shown how to build image mosaics from endoscopic video sequences and applied it successfully for the first time to color images from neuroendoscopic interventions. It has shown to be fast enough to run as a side task in the operating room.