Automatic feature generation in endoscopic images

Motivation: Fiber optic endoscopy is essential for minimally invasive surgery, but endoscopic images are very challenging for computer vision algorithms, since they contain many effects like tissue deformations, specular reflections, smoke, variable illumination and field of view. We developed a method to extract features from endoscopic images usable for scene analysis and classification. These features could be used with data from other sensors for workflow analysis and recognition. Materials and methods: Evolutionary reinforcement learning that automatically computes good features, making it possible to classify endoscopic images into their respective surgical phases. It is especially designed to abstract the relevant information from the highly noisy images automatically. Results: Automatic feature extraction was used to classify images from endoscopic cholecystectomies into their respective surgical phases. These automatically computed features perform better than some classical features from computer vision. The automated feature extraction process enables reasonable classification rates for complex and difficult images where no good features are known. Conclusion: We developed an automatic method that extracts features from images for use in classification. The method was applied to endoscopic images yielding promising results and demonstrating its feasibility under demanding conditions.