The objective of our research work is to control in a visual closed-loop manner the full motion of a 2D (US) ultrasound probe actuated by a medical robot in order to reach and track a desired cross-section image. Considering the ultrasound modality in visual servoing is not an easy task since an US probe differs strongly from a camera. In fact, an US probe provides only information in its observation plane whereas a camera gives a 2D projection of the complete 3D world. Moreover, US images exhibit very low SNR that increases the difficulty to extract in real-time robust visual information. The aim of our research activity is therefore to propose generic solutions for ultrasound visual seroing. We focus more particularly on the modeling of visual features that can be robustly extracted from US image and that are pertinent to visually control the full motion of a robotized probe. We propose two categories of approaches. The first considers geometrical features as input of the control scheme that are computed from image moments extracted after a segmentation step of the peroperative US image. This approach has been used to automatically position the probe on a desired organ cross-section or to perform registration of 3D preoperative data on 2D US peroperative images. The second approach uses directly ultrasound dense information as input of the robot controller. It eliminates the image segmentation step and allows the consideration of weakly structured images where organ contours are hard to detect. This approach has been applied to automatically compensate physiological motion in order to stabilize the observed peroperative US image.
