Endoscope Distortion Correction in Automated Diagnosis Support Systems

Modern endoscopes are using high resolution CCD and storage of pictures and video sequences can be done using additional hardware. Therefore, analysis of images or video sequences with the assistance of computers is possible in principle. This allows computer assisted detection and classification of lesions. A specific type of degradation present in endoscopical imagery is a barrel-type distortion due to the wide-angle or fisheye nature of the endoscopic optics`. While distortion correction in endoscopy has been applied for various purposes, the possible benefits for computer-based diagnosis support systems based on mucosa surface classification or retrieval has not been addressed in literature and actual systems so far. Consequently, this project proposal aims at developing techniques for a computer assisted clinical assessment employing and comparing different types of distortion correction techniques and their corresponding interplay with subsequent feature extraction techniques. In particular, we will focus on adapting distortion correction and feature extraction methods to allow for optimal interplay of these two techniques. As an alternative to this approach, we will also consider distortion compensating feature extraction, where based on the analysis of the underlying distortion parameters, scale- and affine-transform adapted features are employed to compensate for the optics` distortion without using explicit classical distortion correction. The developed algorithms will be evaluated using three datasets acquired with different types of endoscopes and with entirely different diagnostic aims (where for all three histopathologic ground truth is available to rate the obtained classification results): First, we use imagery acquired from the duodenum during gastroscopy using classical white light endoscopy aiming for automated diagosis of celiac disease. Second, we use images acquired with a magnification colonoscope using chromo-colonoscopy aiming for staging of colonic polyps. Third, images acquired with a recent HD virtual chromoendoscopy colonoscope are used.

Modern endoscopes are using high resolution CCD and storage of pictures and video sequences can be done using additional hardware. Therefore, analysis of images or video sequences with the assistance of computers is possible in principle. This allows computer assisted detection and classification of lesions. A specific type of degradation present in endoscopical imagery is a barrel-type distortion due to the wide-angle or fisheye nature of the endoscopic optics'. While distortion correction in endoscopy has been applied for various purposes, the possible benefits for computer-based diagnosis support systems based on mucosa surface classification or retrieval has not been addressed in literature and actual systems so far.Consequently, this project was aiming to develop techniques for a computer assisted clinical assessment employing and comparing different types of distortion correction techniques and their corresponding interplay with subsequent feature extraction techniques. In particular, we have focused on adapting distortion correction and feature extraction methods to allow for optimal interplay of these two techniques. As an alternative to this approach, we have also considered distortion compensating feature extraction, where based on the analysis of the underlying distortion parameters, scale- and affine-transform adapted features are employed to compensate for the optics' distortion without using explicit classical distortion correction.The developed algorithms have been evaluated using imagery acquired from the duodenum during gastroscopy using classical white light endoscopy aiming for automated diagosis of celiac disease. Compared to standard CAD systems ignoring the present optical distortion, we have achieved partially significant improvements wrt. diagnostic accuracy.